YOLOv5与Deep-SORT联合优化的无人机多目标跟踪算法

针对无人机平台下小目标检测性能差、目标尺度变化较大、复杂背景干扰等导致跟踪失败的问题，该文提出一种联合优化检测器YOLOv5（You Only Look Once）和Deep-SORT（Simple Online and Realtime Tracking with a Deep Association Metric）的无人机多目标跟踪算法。该算法使用改进的CSPDarknet53（Cross Stage Paritial Darknet53）骨干网络重新构建检测器中的特征提取模块，同时通过自顶向下和自底向上的双向融合网络设计小目标检测层，采用无人机航拍数据集训练更新优化后的目标检测网络模型，解决小目标检测性能差问题；在跟踪模块中，提出结合时空注意力模块的残差网络作为特征提取网络，加强网络感知微小外观特征及抗干扰的能力，最后采用三元组损失函数加强神经网络区分类内差异的能力。实验结果表明，优化后的目标检测的平均检测精度相比于原始YOLOv5提升了11%，在UAVDT数据集上相较于原始跟踪算法准确率与精度分别提高了13.288%、3.968%，有效减少目标身份切换频次。 

UAV Multi-Target Tracking Algorithm Jointly Optimized by YOLOv5 and Deep-SORT

? ?Aiming at the problems of tracking failure caused by poor detection performance of small targets， large target scale changes， and complex background interference under the unmanned aerial vehicle platform， this paper proposed an unmanned aerial vehicle multi-target tracking algorithm that jointly optimized YOLOv5 （You Only Look Once） and Deep-SORT （Simple Online and Realtime Tracking with a Deep Association Metric）. The algorithm used the improved CSPDarknet53 （Cross Stage Paritial Darknet53） backbone network to reconstruct the feature extraction module in the detector. At the same time， the small target detection layer was designed by the top-down and bottom-up bidirectional fusion network. In the meanwhile， the optimized target detection network model was trained by the unmanned aerial vehicle aerial photography dataset， which solved the problem of poor detection performance of small targets. As for the tracking module， a residual network combined with the spatiotemporal attention module was proposed as a feature extraction network to enhance the network's ability to perceive small appearance features and anti-interference. Finally， the triple loss function was used to strengthen the ability of the neural network to distinguish within-class differences. The experimental results show that the average detection accuracy of the optimized target detection is improved by 11% compared with the original YOLOv5， and the accuracy and precision of the UAVDT（The Unmanned Aerial Vehicle Benchmark: Object Detection and Tracking） data set are improved by 13.288% and 3.968% respectively compared with the original tracking algorithm， effectively reducing the target. Identity switching frequency.