Research On Vision Perception Algorithm Of Autonomous Vehicles Based On Deep Learning

Intellectualization,connectivity,electrification and sharing have become the core direction of development of automobile industry,this became more obvious since the 21 st century,the intellectualization of vehicles has been the key direction and research field for vehicle manufacturers and universities.Because of the advantages of deep learning,neural network in artificial intelligence,autonomous driving perception technology has developed rapidly,and visual perception is an essential part of it.In this paper,the automatic driving visual perception algorithm is studied from three directions: object detection,lane detection and driving free space recognition.This paper first introduces research status of deep learning-based methods of object detection,lane detection and driving free space recognition in China and abroad,then brings up some classic,excellent models,which provides ideas and theoretical basis for the construction of the model in this paper.In second chapter,the network model and the experimental optimization process of the object detection model is introduced.The model is based on Det Net,and we add attention prediction to the channel,fuses and concatenate the extracted three feature images with different depths and adjusts the space and channel based on the attention mechanism.Last,based on the pre generated anchor boxes,the category and position of the prediction boxes are predicted.Softnms is used to handle overlapping targets in the reasoning process,and the training process is the sum of classification loss and regression loss.Focal loss combined with label smoothing technology is used to focus on the samples that are difficult to train to improve the robustness ability of the model.CIo U loss considers the differences in the location,shape and size of the detection frame at the same time.The quantitative assessment of this model on Pascal VOC test set shows that the m AP is 78.3%,and the calculation speed is 28-33fps;the qualitative evaluation of the model is shown in different scenarios.For small objects and traffic participants specific in China,due to the limitations of the data set,the training effect needs to be improved.The third chapter introduces the network model and experimental optimization process of lane detection model.The model is based on the key point detection method,combined with SCNN and Hourglass network to extract the lane features,through two branches to predict the confidence and position offset of lane pixels.The loss function includes the sum of confidence loss and offset loss.The quantitative evaluation of the model on Tusimple test set shows that the acc is 94.23%,and the calculation speed is 23-26fps;the qualitative evaluation of the model is displayed in different weather and road scenarios.Under good weather conditions,the model has good detection effect for clear lane lines,while when the road area is wet,for unclear lane lines,especially for the outside of adjacent lanes,it is easy to miss the inspection of the lane line.The fourth chapter introduces the network model and experimental optimization process of the driving free space recognition model.The model is based on deep lab V3 +,which adopts the encoder decoder structure that are most widely used,a large number of deep separable convolutions are used to reduce the computational overhead,and hole convolution to expand the receptive field.A series of channel attention and spatial attention modules and spatial pyramid pooling module are used to improve the sensitivity of the model to the distribution of input features.The decoding part stitches 4 times down sampled image and 2 times down sampled image.The loss function is added by Dice Loss and CELoss,which not only reflects the regional correlation of detection results,but also improves the stability of training.Quantitative evaluation of the model on cityscape test set shows that the Io U of the driving free space reaches 96.6%.When the lightweight network Mobile Netv2 is adopted,the computational cost can be greatly reduced without much reduction in accuracy.Finally,qualitative evaluation of the model on structured and unstructured roads can better segment the driving road.