基于深度聚合高分辨率网络的多目标跟踪

Multi-object tracking based on deep aggregation high-resolution network

为了提高多目标跟踪性能,在单个网络中结合目标检测和重识别任务的特点,提出了一种基于深度聚合高分辨率网络的多目标跟踪算法.跟踪器通过DLA网络提取抽象的语义特征图,并将其输入到改进的轻量型HRNet网络中,以高分辨率聚合目标的多尺度特征,同时引入重识别分支提高匹配精度.经由传统的相似度计算、运动预测和数据关联阶段,完成跟踪流程.通过消融实验研究了不同融合层组合和特征维度对跟踪性能的影响,并在基准数据集上与当前跟踪器的各项性能指标进行比较.结果表明,所提算法以简洁的主干网络输出高分辨率的深层特征,兼顾了跟踪精度和执行效率.跟踪器的精度和识别率较高,且具备实时跟踪性能.

To improve the performance of multi-object tracking, a multi-object tracking algorithm based on deep aggregation high-resolution network was proposed by combining the characteristics of object detection and re-identification tasks in a single network. The tracker extracted the abstract semantic feature map through deep layer aggregation(DLA) network, and input it into the modified lightweight high resolution network(HRNet) to aggregate the multi-scale features of objects with high resolution. Simultaneously, the re-identification branch was introduced to improve the matching accuracy. The tracking process was completed by traditional similarity calculation, motion prediction and data association stages. The influence of different fusion layer combinations and feature dimensions on tracking performance was studied by ablation experiments, and the performance indexes were compared with those of state-of-the-art trackers on the benchmark datasets. The results show that the proposed algorithm balances the tracking accuracy and execution efficiency with the simple backbone network output of high-resolution deep features. The tracker has high accuracy and recognition rate with real-time tracking performance.