Research On Intelligent Driving Target Detection And Semantic Segmentation Algorithm Based On Deep Learning

With the increase of car ownership,cars have brought convenience to humans,but also caused problems such as traffic jams and frequent traffic accidents.The intelligent of cars has become a key development trend in the domestic and foreign automotive industries.The environment detection technologies such as target detection and semantic segmentation of automobile intelligent driving are the key links.The perception results directly affect the vehicle’s decision-making planning and vehicle execution control.Deep learning methods have made considerable progress in the field of computer vision.Visual perception based on deep learning is a feasible solution for studying intelligent driving.Target detection and semantic segmentation are both core tasks of visual perception.This dissertation mainly studies the intelligent driving target detection and semantic segmentation algorithm based on deep learning,and proposes a joint algorithm model of target segmentation and detection based on deep learning.The specific research contents are as follows:For the research of target detection algorithm based on convolutional neural network,based on the YOLOv3 algorithm,the model structure is optimized and designed,using feature map information that is shallower than the original algorithm;at the same time,the target is generated based on the pixel-level dataset Cityscapes Detect the labeled data set,and then use the clustering algorithm to cluster the target size of the processed data set to obtain 12 preselected boxes for initializing the model.Finally,the optimized model is trained and compared with the original YOLOv3 algorithm,the optimized model improves the detection performance of small targets.Aiming at the research of semantic segmentation algorithm based on convolutional neural network,by comparing the current mainstream semantic segmentation framework,a semantic segmentation model framework for the combination of encoder-decoder and hollow convolution pyramid pooling structure is designed.The encoder is composed of Darknet-53 and the porous space pyramid pooling structure,and the decoder combines the shallow and deep feature maps at the same time.Finally,based on the pixel-level data set Cityscapes,the designed model is trained to achieve the segmentation of 19 types of targets in intelligent driving scenarios.Finally,based on the previous research of target detection and semantic segmentation model,the multi-task task model in this dissertation is proposed,which has two branch tasks of semantic segmentation and target detection.Based on the data set with both pixel-level and target detection formats,Cityscapes trains the multi-task model,and uses the data validation set and test set to evaluate and visualize the model performance.The experimental results show that the branch of the multi-task model has better performance than the single-task model.The multi-task model can simultaneously achieve the semantic segmentation and detection tasks of the target in the traffic street scene,proving that the multi-task model designed in this dissertation has Reliability.