Research On Object And Free Space Detection Of Intelligent Driving Based On Deep Learning

Cars bring convenience and progress to human society,but also cause many social problems such as traffic safety and traffic congestion.Research on new intelligent networked cars is an important means to solve these problems and the intelligentization of automobiles is one of the trends of the development of the future automotive industry.The intelligent driving technology of intelligent vehicles is mainly composed of three parts: perception,decision and control.Among them,reliable perception technology is the prerequisite for efficient decision-making and precise control.Vision-based perception is one of the solutions to achieve low-cost,high-reliability sensing technology for autonomous driving.Object detection and free space detection are two important components.At present,the deep learning method is more and more widely used in the field of computer vision.This paper studies the shortcomings of intelligent driving object detection and free space detection,and designs a deep neural network based on multi-task learning named PercepNet.The main work content is as follows.Firstly,the background and significance of this topic are expounded.The research status of object detection and free space detection in the field of intelligent driving is reviewed.The comparison between the traditional machine learning method and the deep learning method to achieve the detection of object and free space is established.This paper adopts the visual and based on the deep learning method to achieve the detection of object and free space.The demand of intelligent driving for object detection algorithm is analyzed,and the detection algorithm framework based on YOLO v3 is proposed.Firstly,dimension clustering based on open driving video dataset BDD100 K is used to cluster the nine most priority frames for intelligent driving environment,speed up network convergence and improve detection performance.After that,the model structure is improved and shallow layer is utilized.Clear position information of the feature map enhances the small object detection performance of the model.Finally,the improved model is trained and the effectiveness of the improved model is verified by experiments.The classification criteria for the free space are established and the image pixels are divided into a direct drivable area,an optional drivable area,and a non-drivable area.The free space segmentation network model of the coding-decoding architecture is designed.The overall architecture design of the network,the upsampling mode and the decoding module design,and the combination of coding and decoding features are completed.The model is trained based on BDD100 K to realize the detection of the free space in the lane level and the validity of the model is verified by experiments.After that,the actual application requirements of object detection and free space detection are further analyzed.Based on the multi-task learning theory,a multi-task learning network PercepNet that simultaneously realizes object detection and free space detection is proposed based on the hard parameter sharing model.Finised the design of the overall architecture,feature sharing and loss function.Finally,PercepNet is implemented based on TensorFlow framework.PercepNet is trained by BDD100 K object detection and free space detection training dataset and the obtained network model is evaluated in three aspects.The experimental results show that PercepNet's object detection branch and the free space detection branch have good performance and have certain performance improvement compared with the corresponding single-task model.PercepNet's object detection branch and the free space detection branch share a backbone network which reduces the hardware resources required for object detection and free space detection.PercepNet increases the speed of the forward reasoning process to meet the real-time needs of in-vehicle GPUs.BDD100 K contains sample data of different weather,different time periods and different road conditions.Based on BDD100 K training,PercepNet performs well in daytime,night,rainy,snowy days,etc.In summary,this paper proposes a scheme object and free space detection based on deep learning method and carries out related experimental research.This work has certain theoretical reference value for visual perception technology and has certain practical significance for the realization and promotion of automobile intelligentization.