Video Object Detection Based On Convolutional Neural Network

Object detection is a research hotspot and basic task in the field of computer vision.It is a combination of image recognition and object location.Recently,great progress has been made with the help of convolutional neural networks,and object detection techniques have become increasingly sophisticated in different scenarios.Practical applications such as intelligent monitoring,automatic driving,and public security put forward greater demands on object detection technology.However,in the video,due to the influence of object deformation,occlusion,blur,illumination and other natural factors,the existing object detection is difficult to overcome these problems.Therefore,video object detection technology faces more challenges.In the field of video object detection,when using the image-based object detection algorithm to complete the video detection task,in addition to detecting some problems existing in the algorithm itself,there is also lack of timing context information,resulting in inconsistency and independence of the object in time,resulting in features.Insufficient descriptions lead to problems with object missed detection.This thesis improves the defects of the original detection algorithm by studying the object detection method based on convolutional neural network.The main contributions of this thesis are as follows:(1)A video object detection method based on association features is proposed.The feature descriptors containing object motions are constructed by using object change information between frames.The feature description is used to compensate for the lack of context information of single-frame object features.Improve the purpose of the test results.(2)A non-maximum value suppression algorithm based on overlap rate averaging is proposed.By weakening the strong correlation between classification and localization in traditional non-maximum suppression,the independence of classification and location prediction in the network structure is detected,and the final adjustment is made.The parameters of the object position are retained to achieve further correction of the position of the target.(3)This thesis constructs a complete video detection network structure,implements end-to-end training,and omits the steps of data preprocessing between modules.Finally,the method of this thesis is carried out on the Image Net VID dataset.The experimental results verify that the proposed method improves the detection accuracy while maintaining a fast detection speed.