Research On Video Object Detection Based On Feature Fusion And Attention Mechanism

As one of the basic tasks of computer vision,object detection is widely used in many applications,such as video surveillance,smart transportation,and object tracking.The object detection methods based on deep learning could be grouped into two classes: anchor-based method and anchor free method.The anchor-based method makes use of the global context in anchor box to classification and bounding box regression,and achieves great performance in classification.However,as the setting of the anchor boxes relies on prior knowledge,the detectors perform inferior on detecting small and slender objects.The anchor free method frames object detection as a pair-wise key points detecting and matching problem,eliminating the need for designing a set of anchor boxes,and achieves great performance in locating.However,local key points tend to produce false detection.In addition,the above methods based on deep learning all rely on a large amount of labeled data.This limits the application in some scenarios.Take the above mentioned three issues into account,the main research achievements are as follows:First,on the issues of inferior performance on small and slender objects,we proposed an object detection method based on feature fusion to tackle the variable scale problem.By establishing the relationship between the objects,the image scene is introduced.Contextual information assists the detection of small objects and slender objects.Experimental results show that the proposed method can effectively improve the performance of small object detection and improve the locating ability of the model.Secondly,the anchor-free method is prone to mismatch problems due to the complicated image spatial relationship.A video object detection method based on attention mechanism is proposed.By introducing foreground segmentation branches,a global attention map is constructed to suppress background information,so that the model pays more attention to foreground object information and reduces the search space of object detection branches.The experimental results show that the proposed method is beneficial to improving the recall of the model.Thirdly,to solve the problem of deep learning methods relying on a large amount of labeled data,a few shot object detection methods that integrates global semantic information is proposed.By introducing the global feature map pooling structure,the model is strengthened to learn category-independent features,and learns to detect new categories of objects.Experiments show that the proposed method can achieve great performance on the novel classes.Finally,according to the proposed method,a surveillance and UAV video object detection system is constructed,combined with object tracking technology,to realize the functions of automatic labeling,detection and tracking of objects in the video.