Research On Video Semantic Segmentation Technology In Automatic Driving

With the increasing demand for active safety and intelligence of cars in the market,more and more cooperates and scientific research institutions invest in the field of driverless driving to promote the development of driverless system.The environmental perception technologies are the eyes and ears of the driverless vehicle,which provides support for the driverless decision system.In the technology of driverless environmental perception,it is significant to segment the real-time video data collected by vehicle camera quickly and accurately.At present,the researches on the semantic segmentation of driverless scene mainly focus on the direction of only improving the accuracy of semantic segmentation,which cannot meet the requirements of rapid and accurate scene segmentation of driverless.In view of this,this thesis concentrates on the semantic segmentation of video in driverless scene,focusing on improving the speed of semantic segmentation and optimizing the accuracy of semantic segmentation of the model,so as to achieve fast and accurate semantic segmentation of video data in driverless scene.The main contents and innovations of this thesis are as follows.(1)An adaptive video semantic segmentation model based on optical flow feature fusion has been proposed in this thesis.In this model,most of the non-key frames in the video stream obtain high-level semantic features by the traditional complex and slow deep convolution network to the method of optical flow feature fusion through the method of optical flow feature fusion,which saves the calculating time of the model and improves the speed of video stream semantic segmentation.At the same time,aiming at the selection of key frames of video stream,the model makes the selection of key frames of video semantic segmentation change adaptively with the change of video stream through a key frame replacement decision network,which realizes fast and accurate semantic segmentation of video data of the driverless scene.(2)An improved semantic segmentation model which bases on attention mechanism has been proposed in this thesis.For the model designed above,the position dependence and channel dependence of the video frame image in the global perception field are not considered in the process of key frame high-level semantic feature extraction.Based on the idea of attentional mechanism,attentional module is added to the high-level semantic feature extraction process of key frame to capture the position dependence and channel dependence,which enriches the key frame high-level semantic features of frames.Furthermore,the sensitivity of the model to small and similar categories in the data is enhanced,and the segmentation accuracy of the model is improved on the basis of keeping the video semantic segmentation rate unchanged.(3)In this thesis,a driverless video semantic segmentation simulation system is designed and implemented.Based on the video semantic segmentation model proposed in this thesis,the overall framework and functional modules which support the real-time video semantic segmentation system are designed,and the main functional modules involved in the system are implemented.