基于深度学习的无人机载平台多目标检测和跟踪算法研究

Multi Target Detection and Tracking Algorithm for UAV Platform Based on Deep Learning

无人机技术和计算机视觉技术相结合,在民用和军用领域都有着广泛的需求,然而当前算法不能很好的适应无人机视角旋转、障碍物遮挡、目标尺度变化等特殊情况。根据实际的难点和挑战,提出了基于深度学习的无人机载平台多目标检测和跟踪算法。主要工作有:在检测方面,通过公开数据集和实际采集的大量数据,训练了基于Darknet53的检测网络作为检测器;在跟踪方面,使用Car-Reid数据集训练了一个残差网络提取目标外观信息,使用卡尔曼滤波提取目标运动信息,并通过一个融合公式将两个信息进行整合得到成本矩阵,最后由匈牙利匹配算法得到跟踪结果。在UAV123数据集和实测采集数据集上分别进行多组实验验证,得到本算法在视角旋转、目标尺度变化、障碍物遮挡情况下均能进行稳定检测跟踪的结论。

The combination of UAV technology and computer vision technology has a wide range of requirements in the civil and military fields.However,the current algorithms can not adapt to the special conditions of UAV,such as rotation of view angle,obstacle occlusion,target scale change and so on.According to the practical difficulties and challenges,a multi-target detection and tracking algorithm based on deep learning is proposed.The main work is as follows:in the aspect of detection,the detection network based on darknet53 is trained as the detector through the public data set and a large amount of data actually collected;in the aspect of tracking,the car Reid data set is used to train a residual network to extract the appearance information of the target,the Kalman filter is used to extract the motion information of the target,and a fusion formula is used to integrate the two information Finally,the tracking result is obtained by Hungarian matching algorithm.Experiments are carried out on uav123 data set and actual data set respectively,and the conclusion is that the algorithm can detect and track stably under the conditions of rotation of view angle,change of target scale and occlusion of obstacles.