A Targeted Region-based Optical Flow and Motion Vector Deriving Approach for Self-Driving Vehicles using Monocular Vision

In autonomous driving, object dynamics are crucial for improved driving performance and fewer collisions, in conjunction with three-dimensional awareness of the environment. Using an aerial perspective, several previous efforts have followed the course of automobiles and examined optical flow for self-driving transportation. Because they track objects without priority, are susceptible to object obstructions, and experience a lack of situational awareness, these works lack emphasis on regions of interest (RoIs). In this work, we therefore propose to address this research gap by first obtaining the RoIs from the driver's perspective by identifying moving objects, such as vehicles and pedestrians, using an innovative divide and conquer strategy to produce various hierarchical 2D bounding boxes utilizing a segmented image, addressing overlying segments leveraging pixel matching. In order to prioritize regions of interest above other areas and ensure data accuracy and smoothness in the areas of interest, we then created the optical flow for the chosen ROIs by taking into account a scaled energy formulation. Last but not least, two segmented images are used for motion vector generation. To mitigate the effects of object deformation (deploying augmentation), erroneous positives (deploying neighborhood scan), and partial template (deploying template splitting) issues, we conduct the enhanced template standardized cross-correlation within the vicinity of RoIs to assess the movement of entities. The motion vectors, which show the relative velocity of objects in reference to the moving vehicle, are obtained by using the centroids of the old and new objects after template matching. To assess how well the suggested method performs, we choose a dataset from the KITTI dataset and the CARLA simulator. The findings demonstrate the effectiveness of the suggested RoI identification and optical flow determination models, and that motion vectors closely match the actual object relative velocities.