Intersection traffic flow detection based on DSL-YOLOv5+DeepSORT area classification detection

In this paper, a DSL-YOLOv5 + DeepSORT area classification detection method is proposed to run on a vehicle detection board to process intersection videos in real time and obtain traffic information required for traffic signal control in order to promote the intelligence of traffic signal control machines. The L in DSL-YOLOv5 indicates that model sparse training pruning was used to streamline the model and improve detection speed; DS denotes that SIOU Loss is used instead of CIOU Loss, and Soft-NMS is used instead of NMS method to improve the detection performance of the lightened model; The region classification detection method is based on the multi-state features of static pending lanes and dynamic passing lanes existing in the intersection traffic scene, and only the vehicles in the passing detection region of the dynamic passing lanes are target tracked, which significantly reduces the number of vehicles to be tracked by the Deep SORT target, reduces the algorithm redundancy, and further improves the detection speed. The experimental results show that the accuracy of detecting traffic flow and passing state information using DSL-YOLOv5+DeepSORT area classification detection method on RV1126 development board reaches 95.3% and 91.0%, respectively, and the detection speed is increased by 93.8% from the original 16 FPS of YOLOv5+DeepSORT target tracking to 31 FPS, which meets the requirements of the realtime requirements for edge devices to effectively collect intersection traffic information.