Multi-Vehicle Tracking and Counting Framework in Average Daily Traffic Survey Using RT-DETR and ByteTrack

Abstract

The average daily traffic survey is essential for repairing and maintaining road sections. This method is generally conducted using a semi-manual approach that counts vehicles using CCTV. This approach is not effective or efficient because of the potential for human error. This paper have three contributions as follows. First, this paper proposes a framework by applying the RT-DETR architecture for vehicle detection and ByteTrack for vehicle tracking and counting in an average daily traffic survey. Second, this paper proposes a multi-vehicle voting algorithm to filter false identification of the same vehicle during the tracking process prior to vehicle counting. Third, to demonstrate the robustness of the proposed frameworks, we evaluated its performance using seven CCTV camera videos taken from diverse scenes during the day and night. RT-DETR Resnet 101, which is trained in an average daily traffic survey dataset, outperforms all object detection architectures with mAP@50 value of 0.992 and mAP@50-95 value of 0.891. RT-DETR Resnet101 also achieve best F1-score among all object detectors with value of 0.91. This framework using a combination of RT-DETR and ByteTrack also succeeded in counting vehicles with various video backgrounds, with an average counting accuracy value above 83% for all conditions. We compare the effect of using a multi-vehicle voting algorithm with and without showed that counting accuracy increased for each combination with an average increase value of 0.78. RT-DETR also has best performance compared to another object detection methods in detecting vehicles experiencing motion blur, especially in nighttime video scenes. In addition, ByteTrack roles in tracking vehicle objects show its robustness to handle occlusion and vehicle ID switch.

Keywords

• Multi-vehicle tracking counting framework
• average daily traffic survey
• RT-DETR
• ByteTrack
• voting algorithm