Research On Real-Time Target Recognition Technology For Video Compression Domain

The real-time recognition of video targets is one of the hot issues in current research.It is the basic work of video target tracking,video target segmentation and other follow-up studies.However,the existing methods are usually based on video decompression,and the computational complexity is high.It is difficult to achieve real-time target recognition.In view of the above problems,this paper focuses on the target recognition model in the video compression domain.It can effective extract the information data in the video compression domain,thereby improving the recognition efficiency of the target in the video.The main contents of this article are as follows:Firstly,this paper proposes a video motion vector construction method based on image morphology to realize real-time processing of video data.Through the analysis of the video compression process,the method extracts the motion vector data of the macroblock in the video compression domain,and then constructs the motion vector diagram by using image binarization and morphological closure operation.Experiments show that the results of the similarity calculation of the video image foreground image are 0.783 and 7.25 in the cosine similarity and difference value hash method.This rsult is superio to the calculation results of the optical flow method and the background difference method.Secondly,this paper proposes a video space feature vector extraction model-MS-CNN based on convolutional neural network.Based on this,this paper proposes a video timing feature extraction model-MT-LSTM.The MS-CNN model consists of two branches: ResNet network and VGG network.Each branch separately performs spatial feature vector extraction on video data,and then fuses the extracted two feature vectors into multi-dimensional feature vectors.The spatial feature vector output by the MS-CNN is input into the MT-LSTM network for timing feature extraction,and finally the target in the video is identified by the extracted feature data.Finally,the model proposed in this paper is validated on the short-video real-time clsification dataset of the global AI Challenge.The experimental results show that the accuracy of the model can reach 73.4% and the processing speed of the video can reach 26 frames of per second on the dataset.It achieves the frame rate of monitoring video image recording,and realizes the real-time processing effect on video object recognition..In this paper,based on real-time extraction of video motion vector graphics,a target recognition method for video compression domain is proposed.The method can realize the effective recognition of the target without decompressing the video,and provides a new technology for the fields of intelligent transportation,automatic driving and video anomaly detection.