关键词: 智能交通, 慢行交通, 目标检测, 速度提取

Abstract: Speed is an important parameter for non-motorized transportation. The current method of extracting the speed from video through target detection cannot ensure the accuracy of the detection and the stability of the target bounding box simultaneously, and the selected speed calculation datum point (abbreviated as datum point) fluctuates greatly, and thus the acquired speed is inaccurate and unstable. To solve this problem, this paper proposes a fusion detection tracking network and speed calculation method based on YOLOv5(You Only Look Once), which can obtain a more stable and accurate speed. First, we use the target detection and target tracking unit to obtain the bounding box and ID information of the detection target. We can obtain the interest region of the target according to the bounding box and send it to the head detection unit. We then obtain the matching head bounding box by the head detection unit. Secondly, according to the target features in the scene, we determine whether the head detection bounding box belongs to the current target or not. According to the different results, we provide two datum point calculation methods. Finally, the three-dimensional coordinates are obtained by mapping the two-dimensional datum point coordinates, and theresults are substituted into the speed calculation formula to obtain the speed. In addition, we proposed two evaluation indicators, accuracy ( MA ) and stability ( MS ) to quantify the evaluation methods. The detection and tracking effects of the fusion network are verified on the public datasets PETS09-S2L1 and TUD-Stadtmitte, and the effects of the datum point calculation and speed calculation methods are verified on the self-built dual-view collaborative dataset. The experimental results show that the target detection and tracking accuracy (MOTA) of the fusion network is more than 25% higher than that of using a single network. The speed calculation method of this paper is 30% more accurate and 6.28% more stable than the commonly used speed calculation method. In conclusion, the method in this paper can ensure both the detection accuracy and the stability of the target bounding box, and the fluctuation of the selected datum points is less, and the speed obtained by the method in this paper is more accurate and stable.

Key words: intelligent transportation, non-motorized transportation, target detection, speed extraction