Research On Autonomous Driving Scene Recognition Based On Deep Learning

In recent years,the amount of vehicles in China has shown an upward trend year by year,which has also significantly increased road traffic pressure and the danger of driving.Self-driving technology can provide safer travel and is the future development trend of automobiles.The most challenging aspect of self-driving technology is environment perception technology.In this thesis,a camera sensor with rich semantic information and low cost is used to recognize the scene of autonomous driving by using deep neural network.The specific recognition content is divided into lane markings segmentation and three-dimensional object detection.For the lane markings segmentation,the model designed in this thesis first uses the high-precision Dense Net to extract the lane markings features,and then uses the series and parallel dilated convolution layers to obtain multi-scale lane markings information,and then combine the deep and shallow features to improve the accuracy of lane markings edge detection,and finally use the lightweight fully connected network to predict the existence probability of lane markings to prevent misidentification.The proposed model achieved the best results on the CULane dataset.For 3D object detection,this thesis designs a detection model without prior box.The model first uses the high-efficiency Shuffle Net for object feature extraction,then uses the joint pyramid upsampling module to obtain multi-scale information of multi-feature maps,then uses the spatial pyramid pooling module to expand the receptive field of the model,and finally uses six detectors to predict object center position,center offset,2D size,3D size,object depth,and spatial orientation separately.The proposed model achieved excellent performance on the nu Scenes dataset.In addition,the model is combined with a multi-target tracking algorithm to enable it to acquire continuous perceptual information.Finally,in order to verify the generalization ability of the lane markings segmentation model and the three-dimensional object detection model,tests were carried out on the real road.The test results show that the above two models can well detect multiple lane markings in the road and threedimensional information of multiple objects.