为满足视觉跟踪算法对跟踪精度与跟踪速度的要求,提出一种结合目标检测的多尺度相关滤波视觉跟踪算法。所提算法基于深度学习的目标检测算法找出图像中目标的位置和尺寸,利用相关滤波算法对所给出的目标特征进行视觉跟踪,并在多个尺度中搜索最优响应;当检测到相关滤波响应值异常时,停止对模型更新;当连续数帧响应值异常时,则在全图范围内搜索目标位置和尺寸。所提算法通过对跟踪状态进行评估和模型更新率自适应调整,解决了传统相关滤波类算法跟踪误差随时间积累的问题,且具有较大的跟踪速度和较高的精度。结果表明:在Matlab平台下,所提算法的平均定位精度为0.593,平均交叠率精度为0.784,帧率为65.3 frame/s。

In order to satisfy the requirements of visual tracking algorithm on tracking accuracy and speed, a multi-scale correlation filtering visual tracking algorithm combined with target detection is proposed. The proposed algorithm is first used to find the target location and size in the image by the target detection algorithm based on depth learning. The correlation filtering algorithm is then applied to the visual tracking of the given target features and the multi-scale search of the optimal response. When the correlation filtering response appears abnormal, the model stops updating. When the response value of several frames continues to be abnormal, the search of target location and size is then made in the whole image. By the evaluation of tracking states and the adaptive adjustment of model updating rate, the proposed algorithm solves the problem of tracking error accumulation over time in the traditional correlation filter algorithm, and possesses high tracking speed and high precision. The results show that as for the proposed algorithm on the Matlab platform, the average positioning precision is 0.593, the average overlap precision is 0.784, and the frame rate is 65.3 frame/s.