Research On Traffic Scene Intelligent Sensing Technology Based On Audio-visual Fusion

The construction and operation of Intelligent Transportation System can improve the quality of traffic service and the efficiency of traffic facilities,ensure traffic safety and reduce traffic congestion.The traffic scene perception technology is one of the most important technologies in Intelligent Transportation System.Video-based traffic event detection technology has developed a lot of achievements and played a good role in the project,however,its robustness in the night and other special circumstances need to continue to be improved.In fact,a lot of traffic video surveillance projects provide audio information,but the audio data is not fully utilized in traffic incident detection.According to the characteristics of demand for traffic scene intelligent perception,this paper studies the target recognition,event detection and technology integrating video and audio information,which is expected to enhance the intelligent perception ability in traffic scene and enhance the comprehensive control performance of the Intelligent Transportation System.In this paper,we have made a comprehensive and systematic study on the theories related to the existing audio-visual fusion technology.We discuss the recent research progress in Image Processing,Computer Vision,Speech Recognition,Machine Learning and other fields,and focus on these three questions about Sound Recognition,Image Enhancement,Image Recognition,and construct audio-visual fusion traffic scene intelligent perception technology framework.The main work and innovation of this paper are as follows:(1).In order to study the intelligent sensing technology in the night traffic scene better,a traffic video data set and a sound data set were created through the field investigation and data collection.(2).Aiming at the problem of vehicle voice recognition,an intelligent vehicle recognition method based on MLCC feature and DNN model is proposed.The superiority and feasibility of the method are verified by experiments.(3).In order to improve the image at low visibility in the scene,we propose an improved MSRCR algorithm--MSRCR-LR.MSRCR-LR algorithm can greatly improve the standard deviation and mean of the original image and improve the contrast,the algorithm can improve the problem that MSRCR reduced image entropy at the same time,(4)On the issue of image feature extraction,we focus on the deep convolution neural network,and use the CNN neural network model to extract the high-level visual features FC6.The experimental results show that the FC6 layer feature can make the classification results to achieve the best when compared with other visual features.(5).Finally,the three kinds of practical events under the traffic scene are analyzed based on the audio-visual fusion scheme from the point of view of data fusion.In this chapter,we propose an event sensing scheme based on the feature level fusion scheme of audio and visual information and the correct rate of the three most common similar event recognition in night is more than 90%using this scheme.The experimental results show the feasibility and superiority of the scheme.