基于相机模型投影的多目标三维人体跟踪算法

针对基于二维目标检测和卡尔曼滤波的多目标人体跟踪算法在视频拍摄角度不定的情况下,检测算法生成不同角度人体二维检测框的朝向和尺度混淆以及卡尔曼滤波器随机初始化造成的初始跟踪误差逐步放大问题,提出一种基于相机模型投影的多目标三维人体跟踪算法。在人体检测阶段,提出Multi-task RCNN(MTRCNN)网络,使用人体运动趋势指导的三维目标检测替代传统的二维目标检测;通过相机模型在世界坐标系中进行人体检测框的投影。在跟踪阶段,使用目标三维尺度和朝向信息初始化卡尔曼滤波参数,加入三维包围框IOU生成目标匹配分数;通过Kuhn-Munkres(KM)算法进行数据关联。在标注数据集和MOT17数据集上与多种算法相比,该算法具有更稳定的初始跟踪性能,并且有效减少了两种ID Switch错误。

MULTI-OBJECT 3D HUMAN TRACKING BASED ON CAMERA MODEL PROJECTION

Concerning the problem of the 2D bounding box scale and orientation confusion generated by the detection algorithm when detecting the human body from different angles,and the gradual amplification of initial tracking error caused by random initialization of Kalman filter in multi-object human tracking algorithm based on 2D object detection and Kalman filtering,a new multi-object 3D human tracking algorithm based on camera model projection is proposed.In the human detection phase,3D target detection guided by human movement trend replaced the traditional 2D target detection through Multi-task RCNN(MTRCNN).The detected human bounding box was modified by the camera model and projected into the world coordinate system.In the tracking phase,the Kalman filter parameters were initialized by the target’s 3D scale and orientation information.Then 3D Box IOU,movement trend feature and appearance feature were combined to generate a target matching score,and the data was correlated through the Kuhn-Munkres(KM)algorithm.Compared with several algorithms on annotated dataset and MOT17 dataset,the proposed algorithm has more stable initial tracking performance and less ID Switch.