Multi-object Tracking And Crowd Counting In Complex Scenes

In recent years,with the popularization and rapid development of monitoring equipments,growing attention has been recieved to visual tasks in complex scenes.Therefore,it is of theoretical significance and practical value to study computer vision with the increasing amount of video data and the increasingly intelligent shooting equipments.This thesis focuses on the research of multi-object tracking and crowd counting.Furthermore,special algorithm designs for videos shot by UAV are carried out.For multi-object tracking,we propose a novel hierarchical method,which is specially designed for the video sequences shot by UAV and implemented by multiple modules executed in stages.A viewpoint-based subclassification step is added in the object detection stage,and the postprocessing operation filters out the false detection samples.In addition,the frame monitoring stage uses two-norm of inter-frame homography matrix as criterion to judge the intensity of camera movement.On this basis,combined with algorithms such as single target tracking and reidentification,trajectories can be recovered in the data association stage.Our algorithm has achieved excellent evaluation results in Visdrone2019 international competition.For crowd counting,a method catering for single image is firstly studied,and a novel architecture based on scale aggregation network is proposed.The network structure obtains enhanced multi-scale representation with densely connected feature fusion between branches.Features at different scales complement each other,which effectively avoids information redundancy.In addition,a novel loss function incorporating the global spatial structural correlation is introduced.This global structural supervision enforces the network to learn the interactions between pixels in a whole patch,thus breaking the limitation of the extractions for local spatial patterns in existing methods.The performance of the proposed method outperforms state-of-the-art crowd counting approaches according to extensive experiments conducted on the major datasets.The crowd counting method in UAV videos is also studied.We improve the network structure of CSRNet,and exploit low-light image enhancement,semantic segmentation to do additional data processing on the input images and output density maps of the network.The final experimental results demonstrate the effectiveness of the proposed method.