基于改进型目标重识别算法的高速公路车辆轨迹还原

针对高速公路场景下难以实现车辆轨迹精准还原的问题,提出以新近大规模建设的ETC门架系统作为检测载体,将车牌识别与车辆重识别(ReID)技术结合实现更好的轨迹还原效果.高速公路车辆目标多、速度快,常用目标检测算法难以满足属性检测与重识别要求的情况下,对多目标检测与重识别的FairMOT算法结构作出改进,添加多个并行头输出,对车牌、车辆颜色、类型、品牌及重识别等特征同时训练,输出车辆多标签属性;制作数据集并开发一套数据标注工具以满足多目标、多标签及重识别标注需求.实验表明,车牌识别准确率达到97.32％,车辆重识别rank-1为86.94％.采用车牌识别与ReID相结合的方式,车辆轨迹还原准确率超过85％,与ReID及车牌识别单独使用相比分别提高了14.8％与9.89％.该方法以27.5 fps的推理速度在高速公路车辆轨迹还原实践中取得了较好效果,也为智能算法在高速公路领域深化应用提供了实践基础.

Highway vehicle trajectory restoration based on improved object re-identification algorithm

In view of the difficult to recover vehicle trajectory in the expressway, this paper proposed an improved object ReID algorithm based on ETC gantry frame system to enhance vehicle trajectory recovery results. The key innovation of the proposed algorithm was combining license plate recognition and ReID in modified FairMOT network. In particular, it took the license plate, vehicle color, vehicle type, vehicle brand and ReID as the multi-output branch of the training network. The proposed network could quickly extract property and ReID information of numerous rapid objects in expressway compared to previous network. It also developed a labeling software and dataset to deal with multi-object labeling, multi-tab labeling and ReID labeling. Experimental result shows that the proposed algorithm obtains 97.32% accuracy of vehicle license plate recognition, 86.94% rank-1 of vehicle re-identification, and 85% accuracy of vehicle trajectory recovery which is 14.8% and 9.89% higher respectively than the previous methods witch only use ReID or vehicle license plate recognition. In the actual expressway scene, this method can run at 27.5 fps with good vehicle trajectory recovery results. In addition, this work shows that intelligent algorithm has good application prospect in the field of expressway.