针对移动机器人跟踪人体目标时目标因角度大幅变化引起外观改变造成的跟踪无效, 提出了多模板回归加权均值漂移跟踪方法。该方法通过建立目标的多模板模型, 应用均值漂移算法实现目标跟踪。首先, 根据前一帧均值漂移结果和当前帧头肩粗定位结果确定目标模板集, 使其包含目标人体的位姿和角度改变。然后, 采用多模板回归加权均值漂移实现目标的精确定位。在多模板均值漂移中引入回归模型实现颜色纹理特征与目标模型相似度之间的映射, 从而控制模板数量, 保证目标检测的实时性。最后, 分别在视频图像和机器人目标跟踪平台上对所提方法进行实验验证。结果显示, 图像处理平均时间为86.4 s/frame, 满足机器人跟踪的实时性要求。该方法解决了目标特征在跟踪过程中发生变化的问题, 提高了机器人跟踪时对目标人体特征变化的鲁棒性。

To solve the invalid tracking of a human target caused by appearance variations due to large angle change of the target in a robot mobile tracking, a multi-template regression weighted mean-shift algorithm was proposed. The algorithm could implement the target tracking by building a multi-template model of the target and applying mean shift. Firstly, the template set was obtained according to the results from mean shift procedure of the last frame and the coarse location information of head-shoulder model of a current frame, by which the position and angle variation of the target person were included. Then, the multi-template regression weighted mean-shift algorithm was used to determine the precise location of the target person. The regression model was introduced to multi-template mean shift to implement a map from color-texture feature to the similarity of target model to limit the number of templates and to ensure the real-time performance of the target detection. Finally, the proposed algorithm was verified by videos and robot tracking tests. The results show that the image average treatment time is 86.4 ms/frame, which satisfies the requirement of person tracking for a mobile robot. The method solves the appearance variation problem of targets in tracking processing and improves the robustness of human targets to its feature variations.