Multi-object Tracking Research And System Realization Combined With Instance Segmentation

Multi-object tracking technology involves machine learning,statistical analysis,and pattern recognition,which has great research value.Multi-object tracking detects and locates multiple objects in a video sequence,which links objects motion trajectories during the movement and maintains the identity of each object when tracking.For the inaccurate detection,occlusion,and similar appearance features,a multi-object detection and tracking methods is proposed,which is based on instance segmentation and appearance features fusion.In order to achieve the high tracking accuracy,the proposed method focuses on the visible area of the object combined with the spatial information and learns object motion for association through the long and short-term memory network.In this paper,we mainly research on the object detection,similarity measurement,and motion estimation.Our contributions are summarized as follows:1.Multi-object tracking method based on Mask R-CNN and appearance feature fusion.The bounding boxes of adjacent objects are similar and closed,and the deformation can easily lead to the wrong description of the object.In this paper,an instance segmentation method is proposed to enhance the description of the object position with an object mask,which preserves the contour features of the irregular object.A feature extraction branch is designed to extract the deep features,and a novel loss function is formulated to speed up the network convergence.Experiments and results demonstrate that the proposed method can effectively reduce the wrong description of the object.2.Anti-occlusion multi-object tracking method based on spatial attention model.Objects occlusion affects the learning and tracking of the occluded object.In order to improve the tracking accuracy of the occluded objects,an anti-occlusion tracking method is proposed,which enhances the similarities between the objects through learning the similarities of the deep feature with spatial information.The appearance features become more discriminant by training Siamese network to verify the object identity category.To enhance the differentiation of the object features,the proposed method improves the feature fusion based on the spatial attention and preserves the spatial structure information in all feature channels.Experiments and results demonstrate the effectiveness and robustness of the proposed method in the occlusion scenes.3.Deep-feature-based motion estimation method for multi-object tracking.Aiming at the identity switches or tracking drift in the process of object interaction,it is difficult to deal with the motion relationship between multiple trajectories due to the close distance between objects.In this paper,a motion estimation method based on depth feature is proposed,which can track changes in time series based on temporal information.We propose to use the temporal information of long and short time memory network to learn the nonlinear motion model online,predict the object position coordinates,and calculate the motion similarity between objects,which can solve the problem of mismatch between detection object and tracking trajectory.In this paper,we also use motion similarity and appearance similarity to determine the complete trajectory according to the matching degree of track data association,and verify the robustness of the improved algorithm in object nonlinear motion and appearance similar scene.In summary,this paper focuses on object detection,similarity measurement and motion estimation.We combine appearance similarity and motion similarity to achieve data association,improve the success rate of object tracking,association matching degree,and improve the anti-occlusion and stability.Finally,an anti-occlusion and robust multi-object tracking algorithm is implemented in the standard dataset,and an intelligent system is designed to demonstrate the function and performance of the proposed algorithm.