SFFSORT Multi-Object Tracking by Shallow Feature Fusion for Vehicle Counting

Abstract

Standard Multi-Object Tracking (MOT) frameworks are currently categorised into three categories: tracking-by-detection, joint detection, and tracking and attention mechanisms. Infrequently, the latter two frameworks require substantial computing resources. The difficulty of implementing real-time tracking does not apply to vehicle detection at traffic crossings. Not only is it essential to meet real-time requirements for vehicle tracking and detection at traffic intersections, but it is also necessary to address common MOT issues such as target occlusion, repetition technology, error detection, etc. Detection-based tracking has a great deal of potential. This study proposed a shallow feature fusion algorithm based on SORT, called SFFSORT and developed an innovative architecture for vehicle monitoring and counting based on detection tracking. This tracking method is more efficient than both SORT and DeepSORT. It achieved 60.9% MOTA and 65.5% IDF1 in MOT16 while MOTA achieved 60.1% and 64.7% IDF1 in MOT17. Utilizing this tracking method as a foundation, we have developed a vehicle counting framework and successfully implemented it on road traffic videos sourced from the Malaysian transportation department. The tracking algorithm presented here effectively addresses the tracking challenges arising from both detection errors and inaccuracies, providing a robust solution. The experimental findings demonstrate that the deep learning framework is capable of achieving lane-level vehicle counting even in scenarios with limited labelled data.

Keywords

• Intelligent transportation
• multi-object tracking
• video tracking
• deep learning
• transfer learning