Weapon detection, misbehavior detection, and facial recognition-related violence pose a significant threat to public safety, especially in urban areas. Traditional closed-circuit television (CCTV) systems rely on manual monitoring, often leading to delayed responses. This study proposes a real-time gun detection system that uses a deep learning (DL) algorithm and convolutional neural networks (CNNs) for accurate and rapid firearm identification in surveillance footage. The system automates weapon detection, issues instant alerts, and supports timely intervention. Integrating with existing CCTV infrastructure enhances situational awareness and response efficiency, contributing to safer environments through intelligent, automated surveillance driven by advanced computer vision technology [1]. Improving security and stopping criminal activity require close observation of human behavior in public and private areas. The manual observation, a major component of traditional surveillance systems, is laborious and prone to human error. Automated behavior analysis has become more accurate and efficient due to recent DL developments. CNN is one of the examples of DL model that is used in this article to identify and evaluate human behavior in real time. The study aims to enhancing surveillance systems by facilitating the intelligent, real-time identification of anomalous, or questionable activity patterns in intricate settings [2]. Video surveillance has evolved over time, transitioning from analog to digital systems. Due to increased public interest in public safety and crime prevention, image security and surveillance systems have grown significantly, supported by the availability of high-quality, affordable imaging technology [3]. Security is always a top issue in all fields because of the increased crime rate in congested regions and suspicious isolated places. Computer vision has significant uses in abnormal detection and monitoring to address various issues. Owing to the increasing need for safety, security, personal property protection, and intelligence monitoring, video surveillance systems that can recognize and comprehend scenes and unusual events are crucial [4]. Facial expression-based emotion recognition has drawn a lot of interest. People’s widespread feelings of dread and insecurity highlight the urgency of proactively addressing such dangers. Handguns, knives, firearms, and other portable weapons are often used in numerous terrible crimes that endanger public safety [5]. The monitoring system’s original investment and ongoing maintenance costs are substantial. These expenses are mostly related to the required human and hardware resources. The surveillance team is a crucial but sometimes disregarded component of surveillance systems [6]. Although manual monitoring is laborious and prone to human mistakes, surveillance cameras have become essential for maintaining public safety. As artificial intelligence advances, DL methods provide a potent way to automate identifying illegal activity. CNNs are used in various DL models. It is possible to analyze video frames in real time and identify suspicious behaviors accurately. This approach enhances surveillance systems by providing timely alerts, reducing response time, and improving overall security [7]. Real-time surveillance systems have become essential in ensuring public safety, particularly in dynamic and densely populated environments. This study suggests a DL-based method for the spatiotemporal detection of abnormal human behavior. By integrating CNNs, the system effectively captures spatial and temporal patterns in video data [8]. Thus, the development of an automated border surveillance system that can perform the surveillance job without requiring human involvement is urgently needed [9]. CCTV cameras are considered one of the most critical security-related criteria and are an essential part of the solution to this problem. The main goals of CCTVs, which are currently installed in every public space, are security detection, safety, and criminal investigation. CCTV footage is the most crucial evidence in courts [10].
To create a real-time crime monitoring system (CMS) that combines facial recognition, behavior analysis, and weapon detection to identify and react to crimes via video surveillance automatically.
To utilize spatial hierarchies in video frames to effectively identify visible and concealed firearms by using the hyper capsule network (H-CapsNet).
To extract spatiotemporal information from video sequences using a 3D convolutional neural network (C3D) to detect unusual or suspicious behavior.
To reliably identify offenders, compare photographed pictures to a saved database using deep CNNs for facial recognition.
The study presents a comprehensive CMS architecture that combines facial identity, misbehavior recognition, and weapon detection to enable real-time detection and warning in public surveillance systems.
It uses an H-CapsNet to identify both visible and concealed weapons. This network uses the spatial hierarchies in video frames to provide a more reliable and precise identification.
Spatiotemporal data in movies are efficiently captured and analyzed using a C3D to detect aggressive and aberrant human behaviors.
The remaining portion of the study is divided into significant sections, which are described as follows: Section II examines the current research efforts in real-time crime monitoring systems using deep learning for weapon, behavior, and facial detection used by different authors. Section III explains the workflow of the suggested approach in the proposed methodology. Section IV presents the findings, analysis, and performance data. Section V presents the conclusion.
Mukto et al. [11] introduced a novel and efficient CMS designed to leverage video surveillance for real-time crime detection and immediate notification to law enforcement authorities. The system was developed to overcome typical human limitations, such as inattentiveness, delayed responses, and oversight in recognizing criminal behavior. By integrating DL models with advanced image processing techniques, the CMS effectively utilized the capabilities of CCTV networks. The framework operated in three critical stages: facial recognition for suspect identification, violence detection for recognizing aggressive actions, and weapon detection to flag potential threats.
Panda et al. [12] focused on classifying cyberattacks using the recent, noisy, and unbalanced Internet of Things (IoT)-Botnet dataset known as UNSW-NB15 Dataset (UNSW-NB15). They use four contemporary methodologies to validate the ideas that they have suggested. A semi-naive Bayesian averaged two-dependence estimator (A2DE) and Hierarchical Graph Convolution (HGC), which combine genetic algorithms with K-means clustering, enhance the chi-square automated interaction detection decision tree method. Additionally, two DL models are used: a CNN-based classifier and a deep multilayer perceptron. The goal of the project is to enhance the precision and effectiveness of cyberattack detection in intricate IoT systems.
Vuyyuru et al. [13] proposed an inventive response to the rising rates of crime worldwide that is an automated drone-based street crime detection system. The technology analyses drone photos using sophisticated CNN models. Images are broken down into base and detail layers using the embedding bilateral filter technique to improve detection accuracy. The proposed fusion model, Information Retrieval (IR) with attention-based Conv-ViT, successfully combines Inception-V3, ResNet-50, and Convolution Vision Transformer. This integration captures both shape and texture features efficiently, making the system robust for real-time crime detection in dynamic urban environments.
To determine the network systems anomaly, we used CNNs. Akhtar and Fen [14] presented a revolutionary binary and multiclass classification model. The NSL-KDD Dataset (NSLKDD) dataset was utilized in this instance. Our model uses a CNN to perform both binary and multiclass classification. We developed a DL-based DoS detection model for both datasets.
Nazir et al. [15] proposed a method that utilizes YOLOv5 for object detection, where the obtained temporal characteristics are bounding box coordinates. The topic is presented as a time-series classification job, and Deep Sort is used to monitor individuals across a video clip. The modern robust temporal feature magnitude technique, which uses the inflated 3D ConvNet, is contrasted with the suggested method for pre-processing. Experimental evaluations were conducted using the widely recognized UCF crime dataset to assess performance. The method aims to enhance crime detection accuracy by leveraging both spatial and temporal features.
Sahay et al. [16] used DL architectures to suggest a unique method for crime scene video surveillance systems that identify violence in real time. Here, the objective is to collect crime scene footage from real-time surveillance systems and apply the spatiotemporal technique to extract features using a deep reinforcement neural network-based classification algorithm. The characteristics were extracted and classified after the incoming footage was processed and converted into video frames. Its goal is to identify signs of violence and animosity immediately so that anomalies may be differentiated from regular trends.
Ullah et al. [17] presented an intelligent anomaly detection system based on deep features that can operate well in surveillance networks with less temporal complexity. The suggested architecture uses a multi-layer bidirectional long short-term memory model to process a series of video frames after first extracting spatiotemporal information from the frames. This model has been pre-trained with a CNN to enhance its ability to recognize anomalous events. As a result, the system can reliably classify ongoing normal or abnormal activities in complex surveillance environments, making it highly suitable for smart city applications.
Raza et al. [18] developed a predictive system for maternal health issues using artificial neural networks and real-world medical data. They proposed that a Dual-Tree Bidirectional Long-Term Convolutional Network (DT-BiLTCN) is a new DL architecture that combines temporal convolutional networks, decision trees, and a bidirectional long short-term memory network. A total of 1218 samples from hospitals, community clinics, and maternity health facilities were gathered using an IoTs-based risk monitoring system to create the dataset. Using the synthetic minority oversampling technique, the data’s class imbalance is addressed. This approach enhances prediction accuracy and supports proactive maternal healthcare through advanced DL techniques.
Verma et al. [19] reviewed standard methods for detecting and identifying suspicious activities, covering both supervised and unsupervised machine learning (ML) techniques previously used by researchers. These approaches range from modeling individual behavior to analyzing crowded environments and primarily rely on classifiers such as ANN, SVM, and Hidden Markov Models. The study also discusses system models designed to distinguish between normal and abnormal human behaviors, along with various feature selection and detection methods employed in earlier research to improve detection accuracy.
Palanivinayagam et al. [20] focused on vulnerability analysis by extracting key variables, such as time zones, crime areas, and crime frequency, to increase ML algorithm accuracy. Two common datasets were used to evaluate the suggested approaches. The outcomes showed that their feature-building strategy greatly improved the ML models’ performance. The efficiency of the suggested method in crime data analysis was demonstrated by the notable achievement of the maximum accuracy of 97.5% utilizing the Naïve Bayes algorithm on the San Francisco dataset.
Alzahrani and Bamhdi [21] presented a robust approach to identifying botnet assaults on devices connected to the Internet. The CNN model with a lengthy short-term memory processing mechanism was imaginatively integrated with four various types of security cameras to identify two common and important IoT hazards. Real-time lab-connected camera equipment in IoTs settings provided the datasets, which included typical malicious network traffic.
Zahrawi and Shaalan [22] provided research based on real-time object detection systems to automatically detect firearms in video surveillance systems. We provide an early weapon identification framework using cutting-edge and real-time object recognition techniques such as You Only Look Once (YOLO) and single shot multi-box detector (SSD). To use the model in practical applications, we also considered drastically lowering the number of false alerts. The approach works well with inside security cameras in gas stations, supermarkets, banks, shopping centers, and other establishments.
Zhang et al.’s [23] primary contribution is creating a fraud detection system that combines a DL architecture with a sophisticated feature engineering approach based on homogeneity-oriented behavior analysis. To evaluate the efficacy of the suggested framework, we conducted a comparison analysis using an actual dataset from one of China’s biggest commercial banks. The trial’s findings demonstrate the practicality and efficacy of our proposed strategy for identifying credit card fraud.
Akhtaret et al. [24] employed the CNN-LSTM technique, a hybrid DL strategy, to identify botnet assaults. Without the administrator’s permission, malicious software gets installed on a computer system or infiltrates it. Numerous viruses are used by cybercriminals to accomplish their malevolent goals. The increasing quantity of harmful applications has prompted the development of a novel DL system. The system employs Natural Language Processing (NLP) approaches as a baseline, records local spatial correlations by combining CNNs and Long Short-Term Memory (LSTM) neurons, and learns from subsequent long-term dependencies.
Chinedu et al. [25] examined the advancements made over the past 10 years in applying ML models to support the development of clever solutions to reduce the threat posed by cybercrimes. Their study focuses on published resources from reputable databases and adopts an inquisitive perspective. It downplays the alleged benefits of some reportedly clever anti-cybercrime tactics while highlighting their application and potential [25,26,27,28,29,30].
Kuklin et al. [31] proposed a model to enhance the reliability of biometric authentication systems by addressing data drift—a common issue that reduces model accuracy over time. Their framework uses dynamic data correction to maintain stable performance, which is particularly relevant to surveillance systems using facial recognition.
Kurdthongmee and Kurdthongmee [32] developed a fast and accurate pupil detection method using semantic segmentation fine-tuned on a lightweight CNN backbone. This approach improves eye tracking and gaze estimation in real-time applications, offering insights into optimizing lightweight models for surveillance and human behavior analysis.
Abumalloh et al. [33] conducted a text mining and qualitative study to explore how individuals experience security breaches and cyber-attacks. Their findings provide a human-centered perspective on cybersecurity risks, which complements technical monitoring systems like CMS by highlighting the need for user awareness and trust.
Previous studies on crime detection and video surveillance systems have made notable contributions as in Table 1, but suffer from several critical limitations that hinder their practical deployment:
Limited real-time performance: Many traditional models (e.g., CNN-only or YOLO-based systems) suffer from latency issues due to sequential processing, which delays alert generation during critical incidents [11, 22].
Weak multimodal integration: Several existing systems focus on either weapon detection, misbehavior analysis, or facial recognition in isolation, lacking a unified architecture to handle multifaceted crime detection concurrently [15, 16].
Inadequate handling of spatiotemporal features: Models that rely solely on 2D CNNs or classical motion analysis (e.g., optical flow) often fail to detect subtle, evolving human behavior patterns over time [17].
Poor detection of concealed objects: Conventional object detection frameworks like YOLO and SSD perform well on visible weapons but struggle with partially concealed weapons due to limited hierarchical spatial encoding [22].
Summary of Existing studies.
| References No. | Dataset used | Algorithm used | Result achieved |
|---|---|---|---|
| [26] | AWID dataset | ML algorithm | With a detection rate of 99.75% and an accuracy of 99.45%, the model performed better than the others |
| [27] | CICD DoS2019 dataset | DL algorithm | F1-score and accuracy rate >98% |
| [28] | UCF crime dataset | ML algorithm | The average of 98.0%. In the meantime, our PBVAD-MIM approach yielded an average success rate of 80.7% for the tests |
| [29] | CIC-DDoS 2019 dataset | ML and DL algorithms | Obtained a 99.50% accuracy rate with a delay |
| [30] | Historical daily weather dataset | ML and DL algorithms | Accuracy rate of 96.65% and 84.0% |
DL, deep learning; ML, machine learning; PBVAD-MIM, Priority-Based Vulnerability Assessment and Detection – Machine Intelligence Model.
To address these gaps, our proposed CMS integrates multiple advanced DL models within a unified, real-time pipeline:
The H-CapsNet outperforms conventional object detectors by capturing spatial hierarchies and detecting both visible and concealed firearms with high precision.
The C3D effectively captures spatiotemporal patterns from video streams, enabling accurate detection of anomalous or aggressive behavior in dynamic environments.
A CNN-based facial recognition module ensures rapid and accurate identification of individuals using robust feature embeddings and real-time matching with suspect databases.
Unlike fragmented previous approaches, the proposed system combines weapon detection, behavior analysis, and facial recognition within a single framework, allowing seamless and synchronized monitoring.
Extensive testing on the University of Central Florida (UCF) crime dataset shows superior performance metrics (e.g., 93.5% accuracy, 92.2% F1-score) and response times <2.2 s, making it suitable for real-world deployment.
The suggested CMS uses three different methods to identify crimes in real time, as shown in Figure 1. It uses an H-CapsNet to detect firearms by examining video frames and determining which weapons are hidden or visible. Misbehavior can be identified using a C3D, which can locate odd behaviors. Deep CNNs use face recognition to find and identify suspects. The CMS makes efficient real-time crime monitoring and response possible by combining these elements. The system uses supervised learning to differentiate between benign and dangerous activity. Techniques for data augmentation improve the generalization and dependability of models. Users’ situational awareness increases with CMS, enabling prompt emergency reactions. The system’s potential for practical uses is demonstrated through testing and validation on the UCF crime dataset. All things considered, the CMS provides a strong option for improving public safety via clever video surveillance. It is a valuable tool for law enforcement because of its possibilities.
Proposed methodology block diagram.
Weapon detection, misbehavior detection, and face identification are three tasks that may be performed using this dataset, which includes all CMSs in one group and all routine activities in another. Our testing results show that our MIL strategy for weapon detection, misbehavior detection, and face identification generates a significant performance boost compared with the state-of-the-art techniques. We also provide results on detecting abnormal behavior from several recent DL baselines.
Several classes are involved in the task of detecting weapons. The model must not only detect the presence of a weapon but also accurately identify its type. The model is trained using images of various weapons, such as handguns, from a weapons database. This helps determine whether the weapon being detected is deadly or not. CMS prioritizes searching for firearms because it can ultimately determine whether a crime is occurring, a H-CapsNet object identification method for weapons discovery. From model training to application development, users can quickly transition with the help of the H-CapsNet. Objects of interest in an image are predicted and recognized by H-CapsNet promptly and efficiently.
Fframe – Input frame features
(2) {u_i} = {W_i}.X + {b_i} ui – Output vector of the ith capsule in the primary layer
Wi, bi – Weights and bias matrices
X – Flattened or pooled input feature vector
An important part of the capsule network’s routing procedure is the total input to higher capsule. To provide the input for a higher-level capsule, predictions from lower-level capsules must be combined. Each prediction vector
cij – Coupling coefficient between capsule i and j
{\hat u_{j|i}}
After integrating inputs from lower-level capsules, the output of a higher-level capsule signifies the capsule’s final activation. The probability and instantiation characteristics of the identified feature, such as a weapon, are both encoded in this vector.
vj – Output vector of capsule j, encoding both presence and pose
squash(·) – Nonlinear activation ensuring the output vector length is between 0 and 1
The length of vj represents the probability of the detected entity.
The vector length of the output capsule is used by the final weapon detection decision to assess if a weapon is present in a video frame. After processing, a vector is produced by the weapon-class corresponding capsule vweapon, whose length indicates the probability of detection as in algorithm 1
‖vweapon‖ – Magnitude (length) of the weapon capsule output.
δ – Detection threshold (a value between 0 and 1, e.g., 0.5).
wd: – Final weapon detection decision.
Extract low-level features from F8 using CNN
For each primary capsule i:
Compute output vector:
{u_i} = {W_i}.X + {b_i} For each higher-level capsule j:
For each lower-level capsule i:
Compute prediction vector:
{\hat u_{j|i}} = {w_{ij}} \times {u_i} Initialize coupling coefficients cij using routing softmax
For routing iterations:
For each capsule j:
{s_j} = \sum\limits_i {{c_{ij}}{{\hat u}_{j|i}}} Compute capsule output:
Vj = squash\left( {{v_j}} \right) Update cij based on agreement between vj and
{\hat u_{j|i}} Select vweapon from the capsule corresponding to the weapon class
Final Decision:
If ‖vweapon‖ > δ
wd = “Detected”
Else:
wd = “Not Detected”
Identifying and classifying weapons based on a reliable H-CapsNet model are key components of crime detection. Using images annotated to the data is essential for H-CapsNet’s performance because the finer points of its weapon identification processes are determined by H-CapsNet files, which act as careful configuration files for H-CapsNet.
The C3D model is a supervised learning model that uses 3D convolution networks to train enormous amounts of image data. Furthermore, because of its spatiotemporal structure and 3 × 3 × 3 kernel shape, it is quite successful in learning image data. In the past, 2D convolution networks were used to analyze images. A crucial challenge in computer vision is the automated identification of suspicious or aberrant behavior in real-time, whether in social settings or during video surveillance. DL methods, in particular CNNs, which can collect both spatial and temporal characteristics of video input, are among the most successful ways to tackle this issue.
Convolutional 3D networks, or C3D models, are robust supervised learning frameworks that use 3D convolutional operations to evaluate video data. The C3D model extends convolutions into the temporal dimension, allowing it to learn motion patterns and temporal dynamics across successive frames, in contrast to conventional 2D convolutional networks that only analyze spatial information inside a single visual frame. By combining temporal and spatial data into a single DL architecture, the C3D model marks a substantial breakthrough in video-based behavior analysis. For applications requiring automated monitoring, public safety, and anomaly detection, this makes it a top option.
The spatial and temporal structure of the C3D model, which uses 3D Convolution Networks, makes it very effective in learning picture data about misbehavior. Learning misbehavior patterns in CCTV images is a prerequisite for employing such C3D to identify aberrant conduct in the photographs. Figure 2 illustrates how the C3D model learns and identifies anomalous activity in CCTV footage. To begin with, learning requires learning.
C3D architecture. C3D, 3D convolutional neural network.
The input video clip representation is the foundational step in misbehavior detection using a C3D. It involves segmenting a continuous video stream into fixed-length clips, typically comprising TT consecutive frames, denoted as Vt = {F1,F2,…FT}, Fi ∈ ℝH×W×3. Each frame Fi is a color image with dimensions H × W × 3, representing the height, width, and color channels. This sequential structure allows the model to capture both spatial and temporal features, enabling the detection of dynamic and suspicious human behaviors across multiple frames, rather than analyzing isolated still images.
Vt – Sequence of video frames
T – Number of frames
H, W – Height and width
Frame normalization is a crucial preprocessing step in misbehavior detection that enhances model performance by standardizing the pixel intensity values of each video frame. It involves adjusting the raw pixel values of a frame Fi using the formula
μ is the mean.
σ is the standard deviation of pixel intensities.
This procedure lessens the impact of changing lighting, shadows, and camera exposure by guaranteeing that all input frames have comparable statistical characteristics. Normalized frames decrease overfitting and increase convergence speed, which aids in the model’s learning.
Temporal stacking of frames is a method used to organize a sequence of video frames into a structured format suitable for input into a C3D. In this process, consecutive frames F1,F2,…,FTs are stacked along a new temporal dimension, forming a 4D tensor Vstacked ∈ ℝT×H×W×3. This structure preserves both spatial (height and width) and temporal (frame sequence) information. By stacking frames temporally, the network can analyze motion patterns and dynamic behaviors across time, which is essential for detecting misbehavior and recognizing actions in video surveillance footage.
b is the bias.
3D convolution operation is a key process in misbehavior detection using C3D. Unlike 2D convolution, which only captures spatial features, 3D convolution processes both spatial and temporal dimensions simultaneously. It involves sliding a 3D kernel K over the input video tensor Vstacked, computing features across width, height, and time, producing an output feature map.
The input image in a face recognition system refers to the facial photograph captured by a camera or extracted from a video frame. It is typically represented as a color image in RGB format with dimensions H × W × 3H times, where HH and WW denote the height and width of the image, and the 3 represents the red, green, and blue color channels. This image serves as the primary data source for subsequent processing steps such as face detection, normalization, and feature extraction. The quality and resolution of the input image significantly influence the recognition accuracy.
If is the input face image with height H and width W.
Convert the RGB image to grayscale to simplify feature extraction if required.
A face detector fdet outputs a bounding box Bf for the detected face region within If.
Pixel values can be normalized by deducting the mean μ and dividing by the standard deviation σ to standardize the input
Pass a normalized face image In through a deep CNN fCNN to extract a high-dimensional feature vector Vf.
Transform Vf to a fixed-length embedding vector E using a fully connected layer with weights We and bias be.
Normalize the embedding E to unit length (L2 normalization) for consistent distance comparison.
The suggested CMS uses video surveillance to detect and report crimes in real-time efficiently. It combines deep CNN-based facial recognition for suspect identification, H-CapsNets for weapon detection, and 3D CNNs for misbehavior recognition. The system uses supervised learning to differentiate between normal and deviant actions, as demonstrated by tests on the UCF crime dataset. Data augmentation approaches improve generalization. CMS helps create safer, more intelligent public surveillance systems by enabling quick emergency responses and enhancing situational awareness.
False positive (FP), false negative (FN), true negative (TN), and true positive (TP) are acronyms that indicate the ratio of photos that are accurately recognized as positive, photos that are mistakenly categorized as negative, photos that are mistakenly labeled as negative, and photos that are mistakenly classified as positive.
CMS must have a low false negative rate (FNR) to guarantee system security. FNR is supplied by
F1-score, confusion matrix, and sensitivity, along with other criteria such as recall, accuracy, specificity, and precision, are used to evaluate the efficacy of the neural-based model system. A brief description of these measurements is given below. Accuracy is the percentage of data instances that are properly categorized overall.
Precision indicates what percentage of identifications are correct
This proportion of correctly identified TPs is known as recall.
The accuracy and recall metrics’ harmonic mean is used to get the F1 score. Two classifiers are compared using it.
The performance analysis of the weapon detection model in Table 1 shows that it is overall very successful. The model’s 93.5% accuracy rate indicates that many cases were correctly identified. With a recall of 91.7%, it effectively detects real weapons, while an accuracy of 92.8% suggests its ability to reduce false alarms. Reliability in handling recall and precision is demonstrated by the model’s balanced F1 score of 92.2%. False positive rate (FPR) and FNR are low, at 3.1% and 2.7%, respectively, indicating its dependability in lowering incorrect classifications. For real-time weapon detection applications, this makes it a dependable choice.
As seen in Figure 3 and Table 2, the performance analysis of the weapon detection system shows excellent results, with 93.5% accuracy, demonstrating reliable weapon recognition in various settings. The model’s 92.8% accuracy indicates that it can reduce false alarms, and its 91.7% recall rate suggests it can detect most real weapons. An F1 score of 92.2% confirms a balanced precision and recall performance. Furthermore, the system’s low FPR and FNR, 3.1% and 2.7%, highlight its resilience and decreased probability of misclassification, making it ideal for security applications.
Performance metrics of the weapon detection system.
Weapon detection performance metrics
| Metric | Value (%) |
|---|---|
| Accuracy | 93.5 |
| Precision | 92.8 |
| Recall | 91.7 |
| F1 score | 92.2 |
| FPR | 3.1 |
| FNR | 2.7 |
FNR, false negative rate; FPR, false positive rate.
The photos in Figure 4 show a computer vision and artificial intelligence-based weapon detection system. Using a bounding box marked “pistol,” the algorithm detects a pistol being aimed by a human in the first image with 89% confidence. The system demonstrates its resilience in the second image by detecting a guy brandishing a knife even if it is upside down. Both situations demonstrate how well the technology can identify potentially harmful items like knives and firearms in real time. These kinds of systems are essential for public safety and monitoring because they allow law enforcement to react swiftly to any threats in delicate locations such as shopping centers, schools, or public areas.
Weapon detection.
Figure 5 shows the model’s training accuracy evolution during 31 epochs. The accuracy starts at approximately 54% and improves markedly during the first several epochs, which indicates learning is taking place. Fluctuations are observed, but the norm is upwards after the fifth epoch, where it levels off at 85%. In the later epochs, the accuracy improves >90%, indicating good fine-tuning of parameters. The gradual rise and plateau suggest minimal overfitting, and the model converges during training.
Accuracy plot.
Figure 6 displays the training loss across 32 epochs. The graph is declining. Initially, the loss is high (>0.7), but it gradually decreases throughout the epochs. After epoch 10, there are a few minor oscillations, but generally, it is declining and leveling off around 0.2. Since the loss is not growing, the model is convergent and not overfitting the architecture and training function well, as evidenced by the diminishing loss.
Loss plot comparison.
Figures 7A and 7B show the CNN-based face recognition detection model’s accuracy and loss curves, emphasizing how well it learns from photos of firearms and crime scenes. By epoch 10, training and validation losses were decreased to 0.33 and 0.4, respectively, with little overfitting from their initial high values of 0.8 and 0.75. Strong generalization was demonstrated by the accuracy improvements from 85.36% to 91.49% (training) and 83.65% to 91.38% (validation).
(A) Training and validation accuracy and (B) training and validation loss.
A facial recognition system in operation is shown in Figure 8. The same person, identified as “Mukto,” is detected and recognized by the algorithm in four distinct frames. Every detection has a green bounding box around it, and the recognition accuracy is shown by the confidence percentage. The results, which vary from 69.70% to 87.46%, demonstrate that the system operates dependably even when head angle and illumination circumstances change. This illustrates how facial recognition algorithms can reliably and instantly identify people. These systems are commonly used to automate very accurate identification verification in surveillance, security, and attendance monitoring applications.
Face recognition.
To further evaluate the classification performance of the CMS components, confusion matrices were generated for both the weapon detection and misbehavior detection modules. These matrices help illustrate how well the models distinguish between TPs, TNs, FPs, and FNs. For weapon detection, the confusion matrix shows the system’s ability to differentiate between the presence and absence of firearms in real-time surveillance frames. Likewise, the misbehavior detection matrix evaluates the accuracy in identifying suspicious human actions. As shown in Figure 9, the high number of TPs and TNs, along with low misclassification rates, confirms the robustness and reliability of both subsystems.
Confusion matrices for (A) weapon detection and (B) misbehavior detection modules. The diagonal values indicate correct classifications, demonstrating the model’s effectiveness in both safety-critical scenarios.
To evaluate the individual contribution of each module within the CMS, an ablation study was conducted as in Table 3. The study involved selectively disabling each component—H-CapsNet (for weapon detection), C3D (for misbehavior detection), and CNN (for facial recognition)—and observing the resulting system performance on the UCF crime dataset.
Ablation study
| Configuration | Accuracy (%) | Precision (%) | Recall (%) | F1-score (%) |
|---|---|---|---|---|
| Full CMS (H-CapsNet + C3D + CNN) | 93.5 | 92.8 | 91.7 | 92.2 |
| Without H-CapsNet | 88.1 | 87.4 | 85.9 | 86.6 |
| Without C3D | 89.3 | 88.7 | 86.2 | 87.4 |
| Without CNN (face recognition) | 90.2 | 89.8 | 88.1 | 88.9 |
| Only H-CapsNet | 86.7 | 85.3 | 84.5 | 84.9 |
| Only C3D | 84.9 | 83.6 | 82.4 | 83.0 |
| Only CNN | 85.5 | 84.8 | 83.3 | 84.0 |
C3D, 3D convolutional neural network; CMS, crime monitoring system; CNN, convolutional neural network; H-CapsNet, hyper capsule network.
The full CMS configuration outperforms all others, indicating that each module contributes significantly to overall system performance. The largest performance drop is observed when H-CapsNet is removed, highlighting the importance of reliable weapon detection in public safety scenarios. The C3D and CNN modules also show notable impacts, confirming that behavior analysis and face recognition are essential for a comprehensive CMS.
While facial recognition enhances the effectiveness of crime monitoring, it also raises important ethical issues that must be addressed:
Facial recognition systems can exhibit bias against certain demographics, such as people of color, women, or younger and older individuals. This may lead to misidentification or false accusations. To reduce this risk:
We trained the system on diverse datasets.
We plan to include fairness checks and ongoing evaluations to monitor bias.
Continuous surveillance and face matching may infringe on individual privacy, especially in public spaces. To address this:
Facial data is processed locally, and sensitive information is not stored permanently.
The system is intended for use only by authorized law enforcement under regulated conditions.
There is a risk of the system being misused for unauthorized tracking, profiling, or surveillance without proper oversight. Therefore:
We recommend strict legal frameworks and transparent governance policies.
System access should be audited and limited to approve personnel.
The UCF crime dataset, which includes actual surveillance footage of a range of criminal activity and sophisticated DL algorithms, is used in this study to create an efficient and intelligent CMS. The CMS offers a complete solution for real-time crime detection and alerting by combining modules like Deep CNNs for facial recognition, C3D for misbehavior detection, and H-CapsNet for weapon identification. Further improving model generality and resilience is the use of data augmentation techniques like as rotation, scaling, flipping, and panning. With 93.5% accuracy, 92.8% precision, 91.7% recall, and a 92.2% F1-score, the system performs well and has low rates of FPs and FNs. These findings demonstrate how proactive monitoring and quick reaction systems in the planned CMS might greatly enhance public safety. For large-scale validation, future research can concentrate on growing the dataset, adding more contextual variables, and implementing the system in actual smart city settings.