Research On Pedestrian Detection Algorithm In The Field Of Autonomous Driving Based On Convolutional Neural Network

With the development of artificial intelligence,automatic driving in the automotive field has been widely concerned by academia and industry,and pedestrian detection is an important part of the environmental perception of automatic driving.The mainstream R-CNN algorithm,Fast R-CNN algorithm and Faster R-CNN based on convolutional neural network were analyzed.The algorithms have the problem of slow detection speed.The dissertation aims to improve the detection speed of the algorithm under the premise of ensuring the accuracy of the pedestrian detection algorithm,so that it can meet the real-time requirements of pedestrian detection in autonomous driving.The U-NET algorithm suitable for medical cell segmentation is applied to pedestrian detection.Because the algorithm has high detection accuracy in the field of cell segmentation,the detection speed is not good.Therefore,on the basis of the U-NET network model,multi-scale fusion technology is used,the number of feature channels of the convolutional feature map at the end of the encoding part in the U-NET network framework is adjusted,and the input pictures are processed by the batch normalization,which get the improved U-NET algorithm.All pedestrian samples are extracted from the COCO data set through algorithms,and then the training set and the testing set are made as the data set for training the U-NET network and the improved U-NET network.Based on the TensorFlow deep learning framework and Python language programming,the network model is trained through the training set in GPU mode,and the accuracy of the model is verified through the testing set.In order to verify the accuracy and rapidity of the improved U-NET algorithm,the experimental result shows that the Dice coefficient of the classic U-NET network model on the training set is 0.9692,the accuracy of the system is 96.92%,and the time to detect a single static photo is about 1 second.In the testing set,the accuracy of detection also reached 94.60% The Dice coefficient of the improved U-NET network model on the training set is 0.9538,the accuracy of the system is 95.38%,and the detection time of a single static photo is about 0.4 second,and sometimes the detection speed reaches 0.1 second.In the testing set,the accuracy of detection also reached 92.81%.Therefore,the improved network model can reduce the system detection time on the premise of ensuring the detection accuracy,so the improved detection algorithm can be applied to the recognition of pedestrians in autonomous driving.