Research On Vehicle Detection Method Based On Video Compression Domain

Nowadays,deep learning is widely used in the field of target detection.Vehicle detection methods based on deep learning have the advantages of high accuracy,no need for manual feature extraction,and good robustness.However,the current deep learning detection methods are mainly based on the pixel domain,and most surveillance videos are transmitted or stored on the network in the form of video compressed streams.The Discrete Cosine Transform(DCT)coefficients and motion vectors contained in the compressed code stream video have rich texture features and motion information.The detection of vehicles on the basis of the video compression domain can improve the detection efficiency.This paper takes the compressed stream video as the object to carry out research.The research contents are as follows:The paper firstly studies the texture features of DCT coefficients.Based on the introduction of discrete cosine transform theory,this paper studies the AC coefficients in the DCT coefficient block.Experiments show that the AC coefficients in different regions correspond to texture features in different directions.Secondly,according to the characteristic that the image information is mainly concentrated in the low-frequency part of the grayscale image,the paper selects the first 24 coefficients in the Y component DCT block as the feature input channel of the model,and draws on the idea of the YOLOv3 target detection model to construct the DCT-YOLO model.The model introduces an attention mechanism to optimize the backbone feature extraction network,improves the position loss function based on the GIOU algorithm,and uses the K-means++clustering method to generate a priori frame.The paper uses the public dataset UA-DETRAC to evaluate the DCT-YOLO model.When the IOU value is set to 0.75,the DCT-YOLO singleclass vehicle detection accuracy rate is 89.06%,the recall rate is 89.61%,the F1 value is 0.89,the m AP is 86.55%,and the multi-category vehicle detection m AP is 87.22%.The detection accuracy of the DCT-YOLO model is comparable to the YOLOv3 target detection model,while the model size of DCT-YOLO is only 1/3 of that of YOLOv3,and the detection speed is 25%faster than that of YOLOv3.Finally,the paper studies the conversion method of DCT blocks of different sizes and the DCT calculation method of moving macroblocks.And based on this,in the H.264 compressed video,the extraction and conversion of DCT coefficients are realized,and the vehicle detection of compressed video is realized by combining the DCT-YOLO model.The experimental results show that the vehicle detection method based on video compression domain proposed in this paper is feasible.The vehicle detection model constructed in this paper has the advantages of high detection accuracy,lightweight model and fast detection speed,which provides a new idea for edge computing and massive image retrieval of video vehicle detection.