基于路侧摄像头的多目标跟踪算法优化设计

Optimization Design of Multi-object Tracking Algorithm Based on Roadside Cameras

针对当前多目标追踪算法应对路侧交通场景的缺陷,提出一种基于路侧摄像头的多目标追踪算法。选择one‐shot追踪算法路线,基于FairMOT设计神经网络,使单个网络同时生成目标检测结果与外观特征结果,增强实时性效果;采用新的数据关联方式,减少遮挡对追踪器的影响;引入新的运动相似度度量方式——缓冲交并比,弥补线性运动预测模型产生的误差;提出基于速度判别的丢失轨迹移除算法和基于历史位置匹配算法,实现长时间遮挡轨迹的身份恢复。在UA‐DETRAC公开多目标追踪数据集上进行实验,验证该算法有效性。为证明该算法在真实路侧环境的适用性,在国家智能网联汽车(上海)试点示范区开放道路采集真实路侧场景数据。最后,将该算法和SORT、DeepSORT、ByteTrack、FairMOT算法在真实路侧场景数据上进行对比实验。实验结果表明,本算法在identification F‐Score、ID switch、fragmentation、mostly tracked、mostly lost、multiple object tracking accuracy等评估指标上优于其他算法。

In response to the limitations of current multiobject tracking algorithms in handling roadside traffic scenarios,a multi-object tracking algorithm based on roadside cameras was proposed in this paper.First,the one-shot tracking algorithm chosen and a neural network based on FairMOT was built to simultaneously generate both object detection results and appearance feature results,thereby enhancing the real-time performance of the algorithm.Then,a novel data association method was adopted to lessen the effect of occlusion on the tracker.After that,a new motion similarity measurement called buffered intersection over union was introduced to compensate for the errors caused by linear motion prediction models.Subsequently,a velocity-based discriminative algorithm for removing lost trajectories and a history-based position matching algorithm to retrieve the identities of occluded trajectories over lengthy periods of time were developed.Experiments were conducted on the UA-DETRAC public multi-object tracking dataset to verify the effectiveness of the algorithm.Additionally,to demonstrate the applicability of our algorithm in realworld roadside environments,real roadside scene data were collected on open field in the National Intelligent Connected Vehicle(Shanghai)Pilot Demonstration Zone.Finally,comparative experiments between the algorithm proposed and SORT,DeepSORT,ByteTrack and FairMOT algorithms were conducted using real-world roadside scene data.The experimental findings indicate that the proposed algorithm performs better than other algorithms in terms of identification F-score,ID switch,fragmentation,mostly tracked,mostly lost,and multiple object tracking accuracy.