A novel texture-based handcraft feature: spatial aggregated LD2oDP-VLAD for gait-based person recognition

Ref.	Year	Feature used	Dataset used	Type of representation	Covariates	Key contribution	Finding
[16]	2017	MHI, BSIF, HOG	OU-ISIR D, CASIA-B, CMU MoBo	Skeleton-based motion features	Walking speed variations	Proposed a hybrid handcrafted descriptor pipeline that combining motion template (MHI), texture encoding (BSIF), and gradient information (HOG).	The fusion of motion, texture, and gradient descriptors was more effective than using any single handcrafted descriptor alone.
[17]	2018	Motion-based local descriptors + Fisher Vector encoding	TUM GAID and CASIA-A datasets	Motion-based (spatio-temporal)	-	Proposed a spatio-temporal gait recognition algorithm combining handcrafted motion features, Fisher Vector encoding, and linear SVM classification.	Showed that combining Fisher Vector encoding with motion-based descriptors improved gait recognition robustness and discriminative power.
[18]	2020	HTG	OU-ISIR D, CASIA-B, CMU MoBo	Silhouette based, Shape/texture representation	-	Proposed a new descriptor: HTG that accumulates temporal pixel gradients instead of just spatial gradients (unlike HOG).	The temporal gradient encoding improved robustness by incorporating dynamic motion cues in addition to static shape cues.
[19]	2021	SIFT, SURF, Shi-Tomasi edge corner detector	CASIA-A	Silhouette-based, shape/texture representation	General variations	Evaluation of individual and combined feature descriptors (SIFT, SURF, Shi-Tomasi) with multiple classifiers for gait recognition.	Combining descriptors improved recognition accuracy, and Random Forest gives the highest accuracy on CASIA-A.
[20]	2021	PVTM	CASIA-B	Silhouette-based, shape	View variation, carrying bag, clothing	Proposed PVTM that transforms semantically selected gait parts across views to mitigate angle variation and covariate factors.	Achieved better recognition accuracy than multi-view baselines on CASIA-B, proving robustness under bag-carrying and clothing variations.
[21]	2022	SURF, HOG, HOF, MBH	KTH, YouTube, and UCF50 datasets	Optical flow-based motion representation	-	Proposed a video representation for motion recognition called MoSURF, which integrates multiple complementary descriptors	Combined motion-based and appearance-based features for more robust recognition in video.
[22]	2022	HOG	CASIA-B	Silhouette-based (GEI representation)	Clothing, Carrying	Proposed GEI + HOG + SVM approach for gait recognition under clothing and carrying covariates	Clothing changes affected recognition less than carrying conditions.
[23]	2022	ISMAEI (derived from active energy image with mid-body masking using Improved-Sobel-Masking operator)	CAISA-B, TUM-GAID	Hybrid: Appearance + Silhouette-based (Shape 27 Motion AEI)	Clothing, carrying	Proposed a two-branch Re-ID framework that fuses deep appearance features (ResNet50) with handcrafted gait features (ISMAEI) to handle covariate changes in long-term person Re-ID.	Achieved better Rank-1 and mAP on CASIA-B and TUM-GAID.
[24]	2023	LR2P descriptor (Local Rotation Robust Pattern)	OU-ISIR B, CASIA-B	GEI based shape and motion representation	Rotation and clothing	Introduced the handcrafted LR2P descriptor for gait recognition, which is rotation- and clothing-invariant	On OUISIR B dataset achieved 86.12% accuracy across different clothing sets. On CASIA-B dataset achieved 92.02% accuracy across 32 tests.
[25]	2025	LFHEI	CASIA-A, CASIA-B	Optical flow-based motion representation	-	Combined handcrafted LFHEI with deep metric learning to build a robust embedding space.	Showed that frontal-view gait recognition was feasible. Achieved 95.0% accuracy on CASIA-A and 89.9% on CASIA-B datasets.
[26]	2025	SIFT, SURF	CAISA-A	Silhouette-based, Shape/texture representation	-	Evaluated SIFT and SURF features for non-intrusive gait recognition using Decision Tree and Random Forest classifiers.	Random Forest achieved higher accuracy (87.80%) on CASIA-A, and showed effectiveness of these feature-classifier combinations.
[27]	2025	Skeletal landmarks + LightGBM	RUET-GAIT	Skeleton-based	Outdoor uncontrolled environment	Proposed OptiGait-LGBM model for skeleton-based gait recognition with low memory usage and minimal computational cost.	Outperformed ensemble methods (Random Forest, CatBoost) in accuracy, memory usage, and training time.

[i] BSIF, binarized statistical image features, GEI, gait energy image; HOG, histogram of oriented gradients, HTG, histogram of temporal gradients, LFHEI, local fuzzy histogram energy image, MBH, motion boundary histograms, MHI, motion history image, PVTM, part view transformation model, SIFT, scale-invariant feature transform, SURF, speeded-up robust features, SVM, support vector machine.

Block diagram of proposed methodology. GEI, gait energy image; LD2oDP, local directional second-order derivative pattern; SP-VLAD, spatial pyramid vector of locally aggregated descriptors; SVM, support vector machine; VQ, vector quantization.

Example calculation of $I_{θ}^{2} (P_{c})$ with θ = 0. The reference pixel *P_c* is highlighted in blue. **(A)** GEI image patch. **(B)** First-order directional derivative response at direction. **(C)** Second-order directional derivative response at direction. **(D)** Generated binary pattern at direction.

Example calculation of $I_{θ}^{2} (P_{c})$ with θ = 45°, 90°, 135°.

GEI representation of a sample subject and corresponding intensity plot. GEI, gait energy image.

LVQP descriptor and corresponding intensity plot. LVQP, local vector quantization pattern.

LD2oDP descriptor with simple VQ and corresponding intensity plot. LD2oDP, local directional second-order derivative pattern; VQ, vector quantization.

Visualization of the SP level 2 with LD2oDP descriptor. LD2oDP, local directional second-order derivative pattern; SP, spatial pyramid.

VLAD residual histograms across SP level-0. SP, spatial pyramid; VLAD, vector of locally aggregated descriptors.

VLAD residual histograms across SP level-1. SP, spatial pyramid; VLAD, vector of locally aggregated descriptors.

VLAD residual histograms across SP level-2. SP, spatial pyramid; VLAD, vector of locally aggregated descriptors.

Combined VLAD residual histograms across SP level-1 and level-2. SP, spatial pyramid; VLAD, vector of locally aggregated descriptors.

Sample GEI image examples from the CASIA-C dataset. GEI, gait energy image.

Table 2:

View-wise performance of the proposed LD2oDP + SP-VLAD method on the CASIA-B dataset in terms of accuracy, precision, recall and F1 score under NW condition

View	Accuracy (%)	Precision (%)	Recall (%)	F1-score (%)
0°	98.20	98.10	97.97	97.72
18°	95.54	95.03	95.97	94.89
36°	87.05	86.22	87.10	84.73
54°	87.44	85.95	87.10	84.52
72°	92.57	93.41	93.15	91.69
90°	93.07	93.65	93.94	92.60
108°	94.20	94.35	94.76	93.36
126°	94.14	94.17	94.72	93.50
144°	88.74	89.05	89.11	86.96
162°	95.05	94.62	95.16	93.82
180°	98.20	96.91	97.98	97.15

[i] LD2oDP, local directional second-order derivative pattern; NW, normal walking; SP-VLAD, spatial pyramid vector of locally aggregated descriptors.

Table 3A:

Comparison of recognition accuracy (%) at lower viewing angles on the CASIA-B dataset

Method (Ref)	Year	Category	0°	18°	36°	54°	72°
Genetic template segmentation [30]	2017	Handcrafted	97.00	99.49	97.50	94.00	88.00
Specialized CNN [31]	2017	Deep learning	79.10	82.52	76.00	81.63	85.20
Deterministic learning [32]	2018	Handcrafted	97.81	98.80	98.05	95.00	92.56
Deep learning framework [33]	2024	Deep learning	96.20	98.70	97.70	97.10	85.30
Frontal-view gait dynamics + deep transfer learning [25]	2025	Hybrid (handcrafted + deep learning)	96.9	-	-	-	-
Part-based Ensemble CNN [34]	2025	Deep learning	88.30	94.60	97.30	94.90	90.80
GaitSet-based [35]	2025	Deep learning	89.8	96.40	97.70	96.30	91.40
Deep learning fusion + MFO (EfficientNetB0 + MobileNetV2) [36]	2025	Deep learning	95.70	91.90	-	-	-
Proposed method (LD2oDP + SP-VLAD)		Handcrafted	98.20	95.97	87.05	87.44	92.57

[i] Highest view-wise accuracy highlighted in bold.

[ii] LD2oDP, local directional second-order derivative pattern; SP-VLAD, spatial pyramid vector of locally aggregated descriptors.

Table 3B:

Comparison of recognition accuracy (%) at higher viewing angles on the CASIA-B dataset

Method (Ref)	Year	Category	90°	108°	126°	144°	162°	180°
Genetic template segmentation [30]	2017	Handcrafted	90.50	89.50	94.50	92.00	91.28	93.94
Specialized CNN [31]	2017	Deep learning	84.46	82.77	79.26	84.70	86.40	82.00
Deterministic learning [32]	2018	Handcrafted	92.80	91.52	93.10	92.73	91.60	94.52
Deep learning framework [33]	2024	Deep learning	88.90	93.20	88.90	93.20	88.20	90.90
Frontal-view gait dynamics + deep transfer learning [25]	2025	Hybrid (handcrafted + deep learning)	-	-	-	-	-	-
Part-based Ensemble CNN [34]	2025	Deep learning	88.50	90.70	93.40	96.10	94.60	85.70
GaitSet-based [35]	2025	Deep learning	88.40	92.80	96.00	97.70	96.10	82.70
Deep learning Fusion + MFO (EfficientNetB0 + MobileNetV2) [36]	2025	Deep learning	82.70	85.30	-	-	91.70	98.90
Proposed method (LD2oDP + SP-VLAD)		Handcrafted	93.07	94.20	94.14	88.74	95.05	98.20

[i] Highest view-wise accuracy highlighted in bold.

[ii] LD2oDP, local directional second-order derivative pattern; SP-VLAD, spatial pyramid vector of locally aggregated descriptors.

Table 4:

Experiments conducted on the CASIA-B gait database to evaluate robustness against variations in viewpoint and clothing

View point	Training dataset	Test dataset	Size of training dataset	Size of test dataset	Proposed method (accuracy in %)
0° (NM)	NM + CL	NM	124 × 5	124 × 2	99.59
0° (CL)	NM + CL	CL	124 × 5	124 × 1	96.75
18°(NM)	NM + CL	NM	124 × 5	124 × 2	97.98
18°(CL)	NM + CL	CL	124 × 5	124 × 1	90.32
36°(NM)	NM + CL	NM	124 × 5	124 × 2	93.95
36°(CL)	NM + CL	CL	124 × 5	124 × 1	87.10
54°(NM)	NM + CL	NM	124 × 5	124 × 2	92.74
54°(CL)	NM + CL	CL	124 × 5	124 × 1	82.11
72°(NM)	NM + CL	NM	124 × 5	124 × 2	97.18
72°(CL)	NM + CL	CL	124 × 5	124 × 1	95.97
90°(NM)	NM + CL	NM	124 × 5	124 × 2	96.77
90°(CL)	NM + CL	CL	124 × 5	124 × 1	95.97
108°(NM)	NM + CL	NM	124 × 5	124 × 2	95.97
108°(CL)	NM + CL	CL	124 × 5	124 × 1	91.94
126°(NM)	NM + CL	NM	124 × 5	124 × 2	95.97
126°(CL)	NM + CL	CL	124 × 5	124 × 1	92.68
144°(NM)	NM + CL	NM	124 × 5	124 × 2	95.16
144°(CL)	NM + CL	CL	124 × 5	124 × 1	95.92
162°(NM)	NM + CL	NM	124 × 5	124 × 2	97.18
162°(CL)	NM + CL	CL	124 × 5	124 × 1	83.99
180°(NM)	NM + CL	NM	124 × 5	124 × 2	99.19
180°(CL)	NM + CL	CL	124 × 5	124 × 1	96.77
Avg (NM)					96.51
Avg (CL)					91.77

Table 5:

Experimental conducted on the CASIA-B gait dataset to evaluate robustness against viewpoint and carrying variations (e.g., walking with a bag)

View point	Training dataset	Test dataset	Size of training dataset	Size of test dataset	Proposed method (accuracy in %)
0° (NM)	NM + BG	NM	124 × 5	124 × 2	99.19
0° (BG)	NM + BG	BG	124 × 5	124 × 1	95.12
18°(NM)	NM + BG	NM	124 × 5	124 × 2	97.98
18°(BG)	NM + BG	BG	124 × 5	124 × 1	89.52
36°(NM)	NM + BG	NM	124 × 5	124 × 2	94.76
36°(BG)	NM + BG	BG	124 × 5	124 × 1	87.10
54°(NM)	NM + BG	NM	124 × 5	124 × 2	93.95
54°(BG)	NM + BG	BG	124 × 5	124 × 1	86.89
72°(NM)	NM + BG	NM	124 × 5	124 × 2	95.97
72°(BG)	NM + BG	BG	124 × 5	124 × 1	95.16
90°(NM)	NM + BG	NM	124 × 5	124 × 2	97.18
90°(BG)	NM + BG	BG	124 × 5	124 × 1	94.35
108°(NM)	NM + BG	NM	124 × 5	124 × 2	95.56
108°(BG)	NM + BG	BG	124 × 5	124 × 1	95.16
126°(NM)	NM + BG	NM	124 × 5	124 × 2	95.56
126°(BG)	NM + BG	BG	124 × 5	124 × 1	93.50
144°(NM)	NM + BG	NM	124 × 5	124 × 2	95.16
144°(BG)	NM + BG	BG	124 × 5	124 × 1	93.55
162°(NM)	NM + BG	NM	124 × 5	124 × 2	97.58
162°(BG)	NM + BG	BG	124 × 5	124 × 1	94.26
180°(NM)	NM + BG	NM	124 × 5	124 × 2	98.79
180°(BG)	NM + BG	BG	124 × 5	124 × 1	97.77
Avg (NM)					96.51
Avg (BG)					92.94

Comparison of the proposed method with [1] on the CASIA-B gait database under viewpoint and clothing variations.

Sample GEI image examples from the CASIA-C dataset. FW, fast walking; GEI, gait energy image; SW, slow walking.

Table 6:

Cross-condition performance of the proposed method on the CASIA-C dataset under different walking conditions

Condition	Accuracy (%)	Precision (%)	Recall (%)	F1-score (%)
FW	81.32	80.08	81.32	78.50
SW	76.37	76.01	76.37	73.31
BW	36.26	35.11	36.26	32.17

[i] FW, fast walking; SW, slow walking.

Table 7:

Comparison of cross-condition gait recognition accuracy (%) on the CASIA-C dataset

Models (Ref)	Year	Category	FW (%)	SW (%)	BW (%)
Local graphical skeleton descriptor [38]	2021	Hybrid (handcrafted + deep learning)	63.6	60.0	42.3
Gait optical flow network [39]	2023	Deep learning	69.3	58.2	43.5
GaitSTAR [37]	2024	Deep learning	71.5	59.1	44.6
Proposed method (LD2oDP + SP-VLAD)		Handcrafted	81.32	76.37	36.26

[i] Top accuracy is in bold.

[ii] FW, fast walking; LD2oDP, local directional second-order derivative pattern; SW, slow walking; SP-VLAD, spatial pyramid vector of locally aggregated descriptors.

Performance analysis of different feature descriptors on the CASIA-B dataset at 0°, 90°, and 180° view angles. LBP, local binary pattern; LVQP, local vector quantization pattern.

Performance evaluation of various feature descriptors on the CASIA-C dataset under NW condition. NW, normal walking.

Table 8:

Ablation experiments on LD2oDP + SP-VLAD with CASIA-B dataset

Method	Codebook size	Stride	SP-levels	0° (%)	90° (%)	180° (%)
LD2oDP + SP-VLAD	64	1	[0, 1, 2]	98.20	93.07	98.20
LD2oDP + SP-VLAD	64	2	[0, 1, 2]	96.85	85.40	97.30
LD2oDP + SP-VLAD	64	1	[0]	80.63	66.29	77.93
LD2oDP + SP-VLAD	64	2	[0]	56.76	38.61	57.66
LD2oDP + SP-VLAD	96	1	[0, 1, 2]	98.30	92.18	98.65
LD2oDP + SP-VLAD	96	2	[0, 1, 2]	98.20	87.71	95.95

[i] LD2oDP, local directional second-order derivative pattern; SP, spatial pyramid; SP-VLAD, spatial pyramid vector of locally aggregated descriptors.

Table 9:

Ablation experiments on LD2oDP + SP-VLAD with CASIA-C dataset

Representation	Codebook size	Stride	SP-levels	FW (%)	SW (%)	BW (%)
LD2oDP + SP-VLAD	64	1	[0, 1, 2]	81.32	76.37	36.26
LD2oDP + SP-VLAD	64	2	[0, 1, 2]	58.24	52.75	27.47
LD2oDP + SP-VLAD	64	1	[0]	32.97	37.91	17.03
LD2oDP + SP-VLAD	64	2	[0]	16.48	15.38	10.44
LD2oDP + SP-VLAD	96	1	[0, 1, 2]	81.32	76.37	36.26
LD2oDP + SP-VLAD	96	2	[0, 1, 2]	58.24	52.75	27.47

[i] FW, fast walking; LD2oDP, local directional second-order derivative pattern; SP, spatial pyramid; SP-VLAD, spatial pyramid vector of locally aggregated descriptors; SW, slow walking.

References

Authors

Metrics

Articles in this issue

DOI: https://doi.org/10.2478/ijssis-2025-0054 | Journal eISSN: 1178-5608

Journal RSS Feed

Language: English

Submitted on: Nov 30, 2025

Published on: Aug 3, 2026

Published by: International Journal on Smart Sensing and Intelligent Systems

In partnership with: Paradigm Publishing Services

Publication frequency: Volume open

Keywords:

gait,

GEI,

Related subjects:

Introductions and overviews,

Engineering, other

© 2026 Ritu Singh, Chiranjit Dutta, Pradeep Kumar Singh, published by International Journal on Smart Sensing and Intelligent Systems
This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.

Volume 19 (2026): Issue 1 (January 2026)

A novel texture-based handcraft feature: spatial aggregated LD2oDP-VLAD for gait-based person recognition

Figures & Tables

Table 1:

Figure 1:

Figure 2:

Figure 3:

Figure 4:

Figure 5:

Figure 6:

Figure 7:

Figure 8:

Figure 9:

Figure 10:

Figure 11:

Figure 12:

Figure 13:

Table 2:

Table 3A:

Table 3B:

Table 4:

Table 5:

Figure 14:

Figure 15:

Table 6:

Table 7:

Figure 16:

Figure 17:

Table 8:

Table 9:

Paradigm

My account