Multi-target Tracking For Unmanned Aerial Vehicle Videos

In recent years,with the development of image processing technology,multi-target tracking based on unmanned aerial vehicle videos has attracted more and more scholars' attention and research.Its application is more and more widely.The tracking feasibility of the algorithm,the robust of the algorithm and real-time are important factors to consider for object tracking.In this paper,we discussed the state-of-art algorithms of multi-target tracking.Aiming at the difficult problems of complex and dynamic background,occlusion and large number of targets,this paper mainly completes the following work:1.Aiming at complex and dynamic background problem of UAV videos,we fully analyze vehicles' space motion information and appearance feature.In this paper,we add the color naming feature and structure motion information as the target feature,and calculate the correlation similarity between target and measurement.First of all,according to the position of measurement,we extraction the color naming feature and structure motion feature of candidates.We use cosine similarity algorithm to calculate the appearance similarity between object and measurement.What's more,similarity of the structure motion feature is calculated based on weighted summation algorithm of combinatorial model.Finally,the product of the two metrics is used as the final similarity measurement.The appearance feature and structure motion information are robust to the dynamic background,which reduce the tracking error caused by irregular motion of UAV and improve the tracking accuracy.2.In order to solve the problem of high complexity and poor real-time performance of the joint probabilistic data association algorithm when the number of target is large,a multi-target tracking algorithm based on m-best data association is proposed in this paper.Target-measurement association probability is mainly determined by the joint events with highest probability.We use the m optimal joint events instead of all events in the traditional algorithm to calculate the marginal probability between target and measurement.In this paper,integer linear programming algorithm is used to solve the optimal joint event,and then the tree-iteration is optimized to find the m joint events.When number of targets is large,the targets are divided into different regions by using K-means clustering algorithm.Then,we use m-best algorithm to associate data within each region and count the remaining amount of target and measured information.The m-best data association algorithm is used to associate the rest of target and measurement.The improved m-best data association multi-target tracking algorithm based on K-means clustering further improves the real-time performance of the tracking algorithm.3.Aiming at serious occlusion and missed detections,we propose a kind of tracklet association algorithm based on Affinity Propagation clustering algorithm.Firstly,the similarity between trajectories is calculated by considering trajectory motion prediction information,trajectory dynamic motion features and color naming features.Then we use the affinity propagation clustering algorithm to cluster the trajectories.Trajectories with the same cluster center belong to the same target.According to the temporal and spatial constraints and inheritance of trajectories,we remove the trajectories which are not qualified.Then trajectory association is carried out.The proposed method takes into account potential motion location information in trajectory interval time and the complexity of trajectory motion.It enriches the description of trajectory motion characteristics.The small trajectory association algorithm provides a uniform identity for the reemerging target,and provides a more complete trajectory feature for target behavior analysis.