A Study Of Deep Learning Based Object Annotation Algorithm For The Surveillance Video Of District Bayonet

It can be found whether there are abnormal people or vehicles entering or leaving the community by detecting and tagging objects in a cell bay surveillance video.On the one hand,the traditional moving object detection based on background modeling can only be used for simple single-object detection but cannot overcome the difficulty of the overlapping conditions.On the other hand,previous deep learning based object detection method only detects the positions and categories of objects from images,regardless of whether it is moving or not.This paper carried out the following two aspects of the work to solve the problems above.An effective background modeling method with similar foreground color and background color is proposed.This paper uses a codebook based multi-element background modeling algorithm,which combines color and texture information,the image of each pixel to create a codebook one by one and constantly updated during the movement of foreground.And we also compared several classic background modeling methods to make sure that the codebook algorithm used in this paper can completely extract the foreground from the entire image.At the same time,this paper proposed a three-channel color histogram equalization to overcome the problem of undetectable large area with similar color and light intensity between foreground and background.And we also use some morphological operations to form a background modeling method with strong anti-interference ability.An effective method for labeling the foreground area of a bayonet movement is proposed.For the problem that the moving objects obtained by background modeling cannot be effectively segmented,the deep learning based object detector is introduced.The moving foreground got by the background modeling is sent to the object detector so that the detector can locates and categorizes each object in each foreground region one by one.Finally,the obtained object class and location are mapped to the whole image.To form a complete and robust cell bayonet video monitoring system for moving target labeling system.Finally,the effectiveness of the algorithm processing is tested through a large number of comparison experiments to verify the effectiveness and robustness of the proposed algorithm.This paper organically combines foreground extraction based on background modeling with object detection based on deep learning.So we can detect and label the moving object in the bayonet monitoring scenes and accurately and effectively monitor and protect people and things entering and leaving the community.