Research On Three-dimensional Multi-Object Trajectory Tracking Of Mobile Robot Based On Instance Segmentation

Moving object tracking is an important research direction in the area of computer vision,which includes obtaining the background area in the image,extracting the moving object area as the foreground,and tracking the targets in image sequences.The dynamic perception information such as the moving object tracking results can be used for behavior analysis and motion prediction,which are helpful in improving the robot positioning accuracy and the degree of navigating intelligence.The traditional methods for this task are generally based on the classical image processing algorithm and the probability models,and experiments showed that these methods have the defects of low accuracy,high noise,and poor robustness.Aiming at the above defects of the traditional methods,this dissertation proposes a novel multi-object segmentation and 3D tracking method called Mask SORT.A CNN-based real-time instance segmentation module is designed for detecting movable objects in image sequences.Based on the segmentation results,LiDAR point clouds,camera poses and the parameters of the multi-sensor system,the 3D coordinates and geometric properties of the objects can be obtained and transformed into the world coordinate system.The 3D coordinate tracker is based on Kalman Filter and data association features extracted by a re-identification feature extracting module,and the motion state of objects can be inferred from the 3D position change.A multi-object segmentation and tracking dataset KITTI MOTS is used to carry out the comparative experiments on Mask SORT and other advanced algorithms designed for the same task.The practical application tests are realized by constructing a sensor platform and collecting data sequences on some complicated scenarios.Experiments show that Mask SORT has high tracking accuracy,strong scene robustness,fast computing speed,and good generalization performance.Mask SORT can effectively deal with some difficult scenes such as occlusion and missing.Compared with the typical realtime instance segmentation networks,the segmenting mAP of the proposed module is improved by 50.2%.The mAP of the proposed re-identification feature extracting module is improved by 30.5%compared with a the classic model,and the Rank-1 is improved by more than 47.1%.As for the multi-object segmentation and tracking performance,Mask SORT improves the sMOSTA metric by 8.8%compared to a typical baseline algorithm,and reduces the processing time by more than 389%.