基于空间特征融合的车辆跟踪算法研究

Vehicle Tracking Algorithm Based on Attention Mechanism and Multi-layer Feature Fusion

跟踪技术是自动驾驶领域非常重要的一种技术。为解决复杂场景下跟踪行驶车辆时产生的目标形变、外观相似等问题,提出了一种结合注意力机制与多层特征融合的多目标跟踪算法。该算法采用ResNe作为主干网络,在特征提取过程中采用特征金字塔网络结构来融合不同网络层级的目标特征。通过在主干网络中添加Coordinate Attention注意力模块来获得更加准确的位置信息,同时在特征金字塔层中使用可变形卷积来适应目标形变的特征提取。在公开数据集UA-DETRAC上进行的对比实验结果表明,改进算法在UA-DETRAC测试集上的多目标跟踪准确率为67.03%,比原ResFPN主干网络提高了4.91%,且跟踪速度为35.87 f/s,能够满足视频帧实时处理的速度需求。因此所提算法具有较好的跟踪效果和实时性。

Tracking is a very important technology in autonomous driving.In order to solve the problems of object deformation and appearance similarity when tracking moving vehicles in complex scenes,a multi-target tracking algorithm combining attention mechanism and multi-layer feature fusion is proposed.Based on ResNet-34 backbone network,feature pyramid network structure is used to integrate the target features of different network levels.In addition,Coordinate Attention module is added in the backbone network to obtain more accurate position information,and deformable convolution is used in the feature pyramid layer to adapt to the feature extraction of target deformation.Results from experiments conducted on UA-DETRAC vehicle pubilc data set show that the improved algorithm has a multi-target tracking accuracy of 67.03%,which is 4.91%higher than the original ResFPN backbone,and the tracking speed is 35.87 f/s.It can meet the requirement of realtime video frame processing speed.Therefore,the algorithm proposed in this paper has good tracking and real-time performance.