基于深度学习的多目标视觉实时跟踪算法研究

Research on multi⁃target visual real⁃time tracking algorithm based on deep learning

文中提出一种基于深度学习的目标检测与目标跟踪相结合的多目标视觉跟踪算法。首先,通过YOLOv3模型进行目标检测,以获得准确的目标检测结果;其次,通过目标检测网络模型依次获取跟踪目标在当前视频帧的候选检测框,并记录坐标位置信息,以获取对应的模板图像及除第1帧外的视频中每一帧图像作为待搜索区域图像;将每个模板图像和待搜索区域图像输入到由孪生卷积神经网络构建的目标跟踪网络框架中,框架对输入的目标模板图像和待搜索图像进行特征提取;最后,通过相关方法计算特征间的相似度,根据相似度得分确定目标的位置与尺度,从而实现连续、实时准确的跟踪。探索成果证明,该算法实时性更佳,与传统的多目标跟踪算法相比较,计算量大大降低,跟踪结果更准确,可以实现对多种目标持续和准确的跟踪,对多目标跟踪具有重要的学术价值和参考价值。

A multi⁃target visual tracking algorithm in combination of the target detection based on deep learning and target tracking is proposed.The YOLOv3 model is used to conduct the target detection to obtain the correct detection result.The candidate detection frames of the tracking target in the current video frame are successively obtained by the target detection network model,the coordinate position information is recorded,the corresponding template image is obtained,and each frame image in the video except the first frame is obtained as an image of the area to be searched.Each template image and the image of the area to be searched are input into the target tracking network framework constructed by twin convolution neural network,and the framework can be used to extract the features of the input target template image and search image.The similarity between features is calculated by cross⁃correlation method,and the position and scale of the target are determined according to the similarity score,so as to realize continuous real⁃time and accurate tracking.The explore results show that the algorithm has better real time ability,and calculation amount is greatly reduced in comparison with the traditional multi⁃target tracking algorithm,the tracking result is more accurate,and can realize the continuous and accurate tracking of multiple targets.It has important academic value and reference value for multi⁃target tracking.