Multiple Object Tracking Algorithm Based On Camera And Lidar

With the development of Artificial Intelligence(AI)and computer vision,Multiple Object Tracking(MOT)has become a research hotspot in recent years.And it has a good application prospect in the fields of intelligent video surveillance,Virtual Reality,medical image processing and autonomous driving.The main task of MOT is to track multiple objects of interest and maintain their identities in a given sequence.However,it may cause tracking failure due to the influence of many factors such as occlusion,similar appearance and scale change.Therefore,how to achieve tracking stability in complex situations is still a rather difficult challenge.In order to solve the above problems,this thesis proposes a MOT algorithm based on camera and Lidar,which combines object detection and object tracking for MOT.The feasibility and reliability of this algorithm are verified through experiments.The main contributions are as follows:(1)Aiming at the problem of low positioning accuracy of tracking object,the object detection algorithm combining image and point cloud is proposed.The improved YOLOv3 algorithm is used in the 2D image.The loss function of YOLOv3 is improved to raise the training convergence speed and detection accuracy,and to solve the imbalance of positive and negative samples.The model is pruned to reduce the parameters,compress the model volume and improve the detection speed.The Point RCNN algorithm is used to process 3D point cloud,and the obtained 3D detection results are projected into the 2D image.Then the object detection results are fused with the improved YOLOv3 algorithm to achieve the complementary advantages of the two detection results.(2)Aiming at the problem of poor object tracking stability in complex environment,an object tracking method based on Deepsort algorithm is proposed.The motion model based on Kalman Filter is established for predicting the locations of the tracking objects at the next moment,and Mahalanobis distance is used for motion matching.And the appearance model based on residual network is established to extract the appearance feature vectors of tracking objects and detection results,and the smallest cosine distance is used for appearance matching.In order to avoid more trajectory fragments and ensure the stability of the tracking trajectories,motion matching and appearance matching are combined for cascading matching to achieve data association and object tracking.(3)A MOT system is built by combining the improved object detection algorithm based on YOLOv3 and Point RCNN with the object tracking algorithm based on Deepsort.The experimental evaluation is performed on the KITTI dataset.The results show that the MOT accuracy can reach 76.68% in the test sequence of MOT benchmark.And the proposed method still keeps good tracking in the case of disordered parking and scale variation of objects.The comparative experiment results in the actual environment show that the proposed MOT algorithm can achieve stable tracking in complex road scenes.