Road Information Detection Method Based On Deep Learning

With the development of artificial intelligence technology,the unmanned and intelligent future travel has gradually become a reality,and the accuracy of road information detection has become the premise and key to the safe and reliable driving of autonomous vehicles.With the continuous improvement of deep learning technology,there have been many research results on road information detection algorithms for common target recognition and semantic segmentation tasks,but most of them focus on a single task and ignore the correlation between road information,and less consideration is given to the computational performance of the vehicle.Therefore,how to use the limited on-board computing resources to realize the simultaneous processing of multiple tasks of object detection and semantic segmentation is the current research focus of autonomous driving technology.In view of the above problems,this topic studies the task of target detection and semantic segmentation in road information detection.By designing a deep learning network model based on multi-task learning,it can process multiple tasks simultaneously without taking up too much computing resources.The main research content is as follows:(1)For the target detection technology,this thesis compares and analyzes the technical characteristics of different algorithm models,builds a target detection network based on YOLOv4,and improves it for road scenes.Firstly,the feature fusion network is improved,and the residual fusion module(Res＿CBL)is designed and used to replace the convolution module in the original structure,which effectively reduces the amount of parameters;secondly,the BDD100 K data set is used to re-cluster the anchor frame that is more suitable for road scenes.In order to reduce the algorithm convergence time and improve the accuracy;finally,the model structure and loss function are improved to improve the accuracy of long-distance small target detection in road scenes.The experimental results show that the improved object detection network in this thesis has good applicability for road scenes.(2)For the research of semantic segmentation technology,this thesis refers to the structure of the encoder and decoder network to build a segmentation network model,and designs specific network modules according to the needs of road scenes.First,in order to ensure the real-time performance of the network,a depthwise separable convolution module(Conv＿ds＿block)with fewer parameters is designed and used in the decoding network.Secondly,in order to make the segmentation results more accurate,this thesis introduces an attention mechanism into the decoding network.Through experimental analysis,the segmentation network designed in this thesis has better segmentation performance under the condition of ensuring real-time performance.(3)In view of the limited on-board computing resources,this thesis refers to the multi-task learning algorithm to build a deep learning network model that can simultaneously complete the tasks of semantic segmentation and object detection.The semantic segmentation algorithm is responsible for the two tasks of lane line detection and drivable area detection.The object detection algorithm Responsible for road vehicle detection.In order to reduce the amount of parameters and make full use of the coupling relationship between different road information detection tasks,the semantic segmentation and object detection networks in the network model share the coding structure and part of the feature fusion structure.Through the experimental verification,the multi-task network model designed in this thesis can complete the detection and segmentation tasks at the same time under the condition of ensuring real-time performance.In summary,according to the task requirements of road information detection,this thesis builds a fusion network of deep learning-based semantic segmentation and target detection algorithms.Experiments show that it can complete the tasks of vehicle detection,drivable area detection,and lane line detection at the same time while ensuring the accuracy and real-time performance of detection.