Complex Traffic Scenes Oriented Free Driving Space Perception Technology Study

With the rapid growth of global car ownership,many cities in the world,especially big cities,are facing such thorny problems as energy waste,excessive exhaust emissions and potential safety hazards caused by traffic jams.Auto-driving has become a powerful technology to improve traffic safety,and it has also become a research hotspot of current frontier science and technology.80%of the external environment information obtained by autonomous cars comes from cameras.Visual-based free-driving space perception technology has become the most important part of automatic driving technology.Image semantic segmentation technology plays an important role in the establishment of free-driving space by labeling the images acquired by the camera pixel-by-pixel into understandable semantic information.Nowadays,the semantic segmentation technology based on deep learning has achieved high segmentation accuracy,but the semantic segmentation algorithm with high segmentation accuracy has the disadvantages of complex model,numerous parameters and large computational complexity.It is difficult to meet the real time requirements of self-driving space perception technology.In this paper,aiming at the problem of large amount of computation in high-precision semantic segmentation network,I analyzed the existing problems of the current algorithm and proposed an improved scheme.First of all,aiming at the problem of large amount of computation and poor real-time performance of high-precision semantic segmentation algorithm.In this paper,the commonly used residual modules in high-precision semantic segmentation model are improved by weight lighting.Combined with convolution grouping and depthwise convolution,an efficient non-bottleneck residual module is proposed,which greatly improves the running speed of the model.This dissertation also improved the spatial pyramid pooling module by weight lighting,which optimized the running speed of the module.Combined with three methods to enhance the segmentation performance of the network model,a lightweight semantic segmentation model is designed.Under the real-time requirement of automatic driving perception technology,the proposed model achieved 65.07%segmentation accuracy,which achieved a good balance between the accuracy and speed of segmentation.Secondly,due to the inevitable multiple use of the down sampling operation of the deep convolution neural network,the location information of the target object in high-level features of the network is seriously lost,resulting in poor segmentation accuracy of the model on the class edge in image,which affects positioning of key targets in driving space.In this paper,we manually add the edge information of the input image to the model,fuse the edge features extracted by the class edge supervised auxiliary branch,and proposed a fusion up sampling scheme,which further improves the segmentation performance of the proposed model of the class edge,and improved the positioning accuracy of autonomous vehicles on key targets in free-driving space.