Enhanced real-time road-vehicles’ detection and tracking for driving assistance

Abstract

Road-object detection, recognition, and tracking are vital tasks that must be performed reliably and accurately by self-driving car systems in order to achieve the automation/autonomy goal. Other vehicles are one of the main objects that the egocar must accurately detect and track on the road. However, deep-learning approaches proved their effectiveness at the expense of very demanding computational power and low throughput. They must be deployed on expensive CPUs and GPUs. Thus, in this work, a lightweight vehicle detection and tracking technique (LWVDT) is suggested to fit low-cost CPUs without sacrificing robustness, speed, or comprehension. The LWVDT is suitable for deployment in both advanced driving assistance systems (ADAS) functions and autonomous-car subsystems. The implementation is a sequence of computer-vision techniques fused together and merged with machine-learning procedures to strengthen each other and streamline execution. The algorithm details and their execution are revealed in detail. The LWVDT processes raw RGB camera pictures to generate vehicle boundary boxes and tracks them from frame to frame. The performance of the proposed pipeline is assessed using real road camera images and video recordings under different circumstances and lighting/shading conditions. Moreover, it is also tested against the well-known KITTI database, achieving an average accuracy of 87%.