Research On The Algorithm Of Object And Lane Inspection Based On Deep Learning

Target detection and lane line recognition are two necessary tasks of the automatic driving perception system,which are the basis for the vehicle to understand the outside world.The detection performance also directly affects the positioning accuracy.Based on the excellent models and algorithms proposed by scholars,this paper establishes target detection and lane line detection models based on deep learning to verify the engineering feasibility of deep learning perception algorithm.Firstly,the forward network part of the target detection model in this paper introduces the algorithm principle of the target detection model in this paper,including the forward prediction process and training optimization process.In the forward network part of the model,based on the classic Res Net50,the cavity convolution module is combined,the feature map size is fixed,and the multi-scale feature fusion is carried out in combination with the PAN structure.In the training part,Focal loss is adopted to focus on the learning of difficult training samples,and combined with the label smoothing technology to increase the model generalization ability.In the position regression branch,CIo U loss is adopted,and the influence of the position,size and shape of the frame is also considered.Finally,the model was simulated on a personal computer,and the model was trained and tuned on PASCAL VOC data set.The effects of data enhancement,number of prior frames,feature extraction network and loss function on the whole model are considered in the experiment.The m AP on the test set is 77%,which is better than the traditional Faster RCNN and SSD,and the detection effect of urban road targets,children and overlapping targets is relatively excellent.Second,the trunk feature extraction network of the carriageway line detection model uses the very lightweight B0 network,which increases the SCNN structure to carry on the fusion to the context semantic information,more effectively detects the carriageway line such a long and narrow target.In the detection part,direct regression fitting line coefficient and predictive confidence were used.The loss function considers the polynomial coefficient regression loss,the interval position regression loss and the classification loss of confidence.The model is trained and tuned on the Tusimple data set.In the experiment,the influence of data enhancement,feature extraction network and polynomial order on the model is considered.The detection accuracy on the test set is 93.77%.Compared with the excellent models such as Lane Net,the prediction accuracy is reduced to some degree,but the prediction speed on this machine reaches 35~37 frames/second,which is greatly improved.Finally,the detection results of the model under severe weather conditions such as rain and snow are presented qualitatively.Finally,based on the algorithm research results in this paper,algorithm experimental research and engineering verification have been carried out for GAC Trumpchi GM8 platform.Considering the complexity of China’s road scene,target detection and the lane line detection algorithm in fault detection/actual road mistakenly identified number should be no more than three times per million kilometers,and the technology to further enhance such as sensory integration in order to get high precision,high reliable results of environmental perception in real vehicle engineering implementation,to ensure the dependability of autonomous perception to the greatest extent.In the next stage,we will continue to study the perception fusion technology based on deep learning and traditional algorithms,and accelerate the engineering application of the algorithm.