Research On Video Surveillance Object Detection Algorithm Based On Prior Knowledge

In recent years,the rapid development of artificial intelligence technology,especially the introduction of deep learning algorithms,has greatly promoted the development of pattern recognition,computer vision and other related fields,and has shown excellent performance in branch tasks such as image classification,object detection,object tracking,instance segmentation,etc.As one of the most challenging branches in the field of computer vision,object detection has also been received wide attention,and has a wide range of application scenarios in real life,such as face recognition,pedestrian detection,consumer electronic,video surveillance,and so on.Object detection aims to identify and locate the target object in the image.The traditional object detection algorithms mostly rely on complex handcrafted feature representation methods and quite a lot of acceleration techniques.With the performance of the object detection algorithm based on traditional manual feature gradually saturated,the development of object detection also appeared a short period of stagnation,until the convolution neural network re-entered people's vision,the object detection based on deep learning theory began to develop rapidly.According to the requirements of the national key research and development plan"Research and development and demonstration of key technologies of intelligent monitoring and early warning and prevention in supervision places",in order to reduce the dependence of a prison video surveillance system on labor and improve the efficiency and reliability of supervision,this thesis studies the automatic detection and recognition of pedestrian targets in prison scenes.Due to the small difference between pedestrian targets in prison video surveillance,the distinguishing characteristics of pedestrian targets are not significant,especially the low image resolution of pedestrian targets under video surveillance,which further increases the difficulty of detection and recognition.On the basis of analyzing and studying the previous work in the field of object detection,based on the prior knowledge of a person in a prison video surveillance,this thesis proposes a video surveillance object detection algorithm based on prior knowledge,the main contributions include:Firstly,for the pedestrian data in a prison video surveillance scene,a pedestrian dataset is constructed manually,and the prior knowledge of pedestrians in the dataset is statistically analyzed,including the resolution of pedestrians in the image,the aspect ratio of pedestrians,the average number of pedestrians in each image and other information,so as to improve the important components of the object detection algorithm;Secondly,an object detection algorithm based on improved feature pyramid network is proposed.In the process of multi-scale feature extraction based on improved feature pyramid network,through a bottom-up enhancement path,the shallow features with high resolution are fused with the deep features of strong semantic information to ensure that the final multi-scale features have strong semantic information and high resolution at the same time,to solve the problem of low resolution of pedestrians in the image;Finally,an object detection algorithm based on the improved non-uniform region proposal network is proposed.The improved non-uniform region proposal network first converts the feature map into a probability map to judge the existence of the target object,and then the prior knowledge of the pedestrian dataset is used to predict the shape of the target candidate box.At the same time,the quality and recall rate of candidate boxes extraction are guaranteed,and the running efficiency of the algorithm is improved.Through the experimental verification,the video surveillance object detection algorithm based on the prior knowledge can achieve higher detection accuracy and faster detection speed in the prison video surveillance scene,ensure the automatic detection and recognition of pedestrians,reduce the dependence of the video surveillance system on labor,and mitigate the workload of the monitoring personnel.