Research And Application Of Small Object Detection And Tracking Algorithm Based On Deep Learning

With the continuous popularization of various land-based,airborne,and space-based monitoring equipment,there is an urgent need for small object detection and tracking in images in fields such as military defense,security tracing,and urban services.Because of the small coverage area,blurry boundary,large displacements and occlusions during movement of small objects,it is difficult for general detection and tracking algorithms to extract the accurate deep feature and matching degree of small objects.In this thesis,aiming at the above challenges,research is conducted based on the algorithm framework of tracking by detection,and improvements will be made from the following three aspects:(1)A backbone network for feature extraction based on contextual information is proposed.Because of the continuous down-sampling operations,the information of small objects will be blurred,lost or even completely disappeared on the deep feature map with small resolution.To solve this problem,a contextual feature extraction backbone network based on swin transformer is proposed to extract the context information in a large receptive field,and then combine it with the feature information extract by the backbone network,to supplement it when features are missing.To further improve the positioning accuracy,a shallow head strategy is proposed,which increases the resolution of the feature map of the detection head to improve the pertinence and applicability of the algorithm for small objects.(2)A multi-path feature fusion network based on a global attention mechanism is proposed.To increase the ability of the network to transfer and fuse object context and feature information,the global attention module is used to amplify the global dimension features,and multiple paths from the shallow layer,deep layer,and backbone network are used to process feature information and context information.Experiments show that on the Vis Drone2019 dataset,the proposed algorithm CGA-YOLO based on contextual information and attention mechanism in this thesis has 5.4% improvement in m AP based on the YOLOv5 l model,and it have more outstanding performance in scenes where small objects are occluded or densely arranged.(3)A multiple objects tracking algorithm based on multi-dimensional information is proposed.In order to solve the problem of tracking and matching failure caused by detection boxes floating due to blurred information of small objects,this thesis improves the algorithm from expanding feature information,increasing positioning difference,and predicting motion trajectory.While increasing the similarity of matching between adjacent frames for the same object,it also improves the tracking and matching rate of small objects in abnormal situations such as occlusion and missed detection.Experiments have shown that on the PESMOD dataset,the proposed multi-dimensional information object tracking algorithm MD-Deep SORT can reduce IDS problems by 42.6% compared to Deep SORT,and has a 9.7% improvement in IDF1 valueLastly,based on the above research,an intelligent traffic density monitoring and object tracking system is designed and implemented.The proposed algorithm is applied to the system,and the testing results show that the algorithm can efficiently complete object monitoring tasks,and has high accuracy and application value.