基于改进YOLOX的移动机器人目标跟随方法

Target Following Method of Mobile Robot Based on Improved YOLOX

针对移动机器人在复杂场景中难以稳定跟随目标的问题,提出基于改进YOLOX的移动机器人目标跟随方法,主要包括目标检测、目标跟踪以及目标跟随三个部分.首先,以YOLOX网络为基础,在其框架下将主干网络采用轻量化网络MobileNetV2X,提高复杂场景中目标检测的实时性.然后,通过改进的卡尔曼滤波器获取目标跟踪状态并采用数据关联进行目标匹配,同时通过深度直方图判定目标发生遮挡后,采用深度概率信息约束及最大后验概率(Maximum a posteriori,MAP)进行匹配跟踪,确保机器人在遮挡情况下稳定跟踪目标.再采用基于视觉伺服控制的目标跟随算法,当跟踪目标丢失时,引入重识别特征主动搜寻目标实现目标跟随.最后,在公开数据集上与具有代表性的目标跟随方法进行了定性和定量实验,同时在真实场景中完成了移动机器人目标跟随实验,实验结果均验证了所提方法具有较好的鲁棒性和实时性.

A target following method of mobile robot based on improved YOLOX is proposed to solve the problem that mobile robots are difficult to follow the target stably in complex scene.This method mainly includes three parts:Target detection,target tracking and target following.Firstly,the lightweight MobileNetV2X network is adopted under the YOLOX framework to improve the real-time performance of target detection in complex scene.Then,the improved Kalman filter is proposed to obtain the tracking state and data association is used for target matching.When the target is judged by depth-histogram,the depth probability constraint and maximum a posteriori(MAP)probability are utilized for matching,which ensure that the robot tracks the target stably under occlusion.Moreover,target-following algorithm based on servo control is proposed,and re-id feature is introduced to actively search for disappeared targets.Finally,qualitative and quantitative experiments on public data set and in realworld environments demonstrate the efficiency of the proposed method.