无人机视角下的多车辆跟踪算法研究

针对无人机视频中存在目标密集、运动噪声强而导致跟踪性能显著下降的问题，提出了一种改进YOLOv3的车辆检测算法及一种基于深度度量学习的多车辆跟踪算法。针对车辆检测的精度与实时性问题，采用深度可分离卷积网络MobileNetv3作为特征提取网络实现网络结构轻量化，同时采用CIoU Loss作为边框损失函数对网络进行训练。为了在多目标跟踪过程中提取到更具判别力的深度特征，提出了一种基于深度度量学习的多车辆跟踪算法，实验证明，本文提出的算法有效改善车辆ID跳变问题，速度上满足无人机交通视频下车辆跟踪的实时性要求，达到17 f/s。

Research on multi-vehicle tracking algorithm from the perspective of UAV

Aiming at the decline of tracking performance suffering from dense targets and strong motion noise in UAV video, we propose a vehicle detection algorithm based on improved YOLOv3 and a multi-vehicle tracking algorithm based on deep metric learning. To improve the vehicle detection system’s accuracy and real-time performance, a deep separable convolution network, MobileNetv3, is adopted as the feature extraction network to realize a lightweight network structure, and the CIoU Loss is used as the frame loss function to train the network. A multi-vehicle tracking algorithm based on depth metric learning is proposed to extract more discriminative deep features during multi-target tracking. Experiments reveal that the algorithm proposed in this paper effectively improves the problem of target ID jump and meets the real-time requirement of vehicle tracking in UAV traffic video, reaching 17 FPS.