基于全卷积网络的多车辆实时跟踪模型

Real-time multi-vehicle tracking model based on fully convolutional network

针对道路多车辆在跟踪过程中由于遮挡或漏检所造成的轨迹ID变换问题,给出一种改进的全卷积网络(fully convolutional network,FCN)和交并比重叠度(intersection-over-union,IoU)数据关联相结合的算法。基于标准卷积和空洞卷积搭建了新的FCN,并进行多尺度目标的定位来增强目标的特征,减少下采样过程的特征丢失;通过在IoU数据关联算法中加入预备跟踪器集合,处理车辆行驶过程中出现的轨迹ID变换问题。实验结果表明,所给出的实时多车辆跟踪算法在UA-DETRAC数据集上具有良好的性能,可以有效降低轨迹ID变换的次数,提高跟踪精度,在实际场景应用中,达到了良好的跟踪效果。

To solve the problem of track ID transformation caused by occlusion or missing detection in the process of multi-vehicle tracking,an improved method combining fully convolutional network(FCN)and intersection-over-union(IoU)data association is proposed.Based on the standard convolution and hole convolution,a new FCN was built,and multi-scale target positioning was carried out to enhance the characteristics of the target and reduce the loss of features during sub-sampled process.Meanwhile,it dealed with the tracking ID switch problem in the process of vehicle driving by adding the set of preparation tracks in the IoU data association algorithm.The experimental results illustrate that the real-time multi-vehicle tracking algorithm has good performance on UA-DETRAC data set,and can effectively reduce ID switch track number.In the actual scene of application,it achieves better tracking effect.