Research On Video Object Detection Based On Neural Network

With the development of computer vision technology and artificial intelligence,the object detection of images has made great progress in recent years,and it has received extensive attention from industry and academia.The object detection of the image needs to detect and identify the specific object,and is also one of the main techniques for video object detection.With the development of intelligent video processing technology,the object detection of video sequences has gradually attracted lots of attention,and become the basis of other technologies in the actual monitoring field.However,how to effectively improve the performance of object detection in video sequences is of great value.To deal with this problem,this paper exploits the key techniques and applications of video object detection from two aspects: the accuracy of image classification in object detection and the correlation between frames in video sequences.This article mainly contains the following research contents:(1)Image classification is one of the core tasks of image object detection.This paper proposes an image classification algorithm based on multi-feature aggregation network.The feature of the image is extracted by using the topology network framework,and then is aggregated and analyzed.The classification performance is therefore improved without increasing the network depth.The aggregation analysis is implemented by a proposed topology network framework based on feedback adjustment learning rate.The topology network consists of multiple topological nodes,each of which is a basic network structure.During the training process,nodes are assigned different learning rates and adjusted at the beginning of each new training period.Firstly,based on the training accuracy of different nodes,the optimal learning rate seed is selected from the current learning rate set.Secondly,a new learning rate set is generated by the learning rate seed and used to train the topology network.(2)In order to improve the accuracy of video sequence detection,this paper uses the correlation between video sequences to estimate the motion vector of moving objects,and proposes an optical-flow-based framework to boost video object detection performance with object enhancement.Firstly,the video frames are grouped to calculate the shared optical flow feature,so that all the video frames in the group share the same optical flow feature.Secondly,the shared optical flow feature is used to enhance foreground information of each video frame in the current group,so that the objects area are highlighted.Finally,the video sequence containing the highlighted area is fed to object detection network.The experimental results show that the proposed method not only improves the accuracy of video sequence detection,but also improves the scene migration performance of the object detection network to some extent.This paper proposes a multi-feature fusion video object detection algorithm based on convolutional neural network.This algorithm improves the classification of object image and make full use of the inter-frame information of video,effectively boosts the performance of video object detection,and has high application value in practice.