Traffic Sign Recognition For Driverless Vehicles Based On DETR

Traffic sign recognition is a key part of automatic driving system and a hot topic in machine vision field for a long time.Traffic signs include a series of important semantic information such as road information,speed limit warning and driving environment,which play a key role in path planning,speed regulation and driving safety of unmanned vehicles.Therefore,real-time and accurate detection of road traffic signs under all-weather conditions is of great significance to the development of autonomous vehicles.The accuracy of traffic sign recognition directly affects the safe driving of automatic driving.However,it is difficult to detect the situation of exposure,reflection and image blur in complex and changeable natural environment.Due to perennial exposure of traffic signs,the surface of some signs appears faded and damaged.The detection performance of the existing algorithms is difficult to reach the expectation.In addition,there are many visionbased road detection tasks for autonomous vehicles,which have high computational requirements.At present,the performance of most algorithms in real-time needs to be improved.Therefore,aiming at the accuracy and real-time performance of the above traffic sign recognition algorithm,this thesis proposes an object detection algorithm based on deep learning,and proposes the following improvement methods for the key parts of the detection task:1.Local image enhancement methods are carried out for the detected traffic signs to improve the contrast and saturation of the area where the traffic signs are located,so that the traffic signs can form a sharp contrast with the image background.The region of interest(ROI)is optimized in the detection process to lay the foundation for subsequent recognition.2.Introduce Attention mechanism into the target detection task of machine vision.The introduction of Attention mechanism can be used to emphasize or select the important information of the target object to be processed in the process of neural network prediction,and suppress irrelevant information.In this thesis,the end-to-end target Detection Transformer(DETR)algorithm is applied to the traffic sign recognition task.The advantage of this method is that it eliminates the post-processing process of non-maximum Suppression(NMS).The prediction performance is improved and the computation overhead is greatly reduced.3.The main network used for feature extraction in DETR model was replaced by Mobile Netv2 network constructed by deep separable convolution,which improved the realtime inference speed of the algorithm,combined with channel pruning and layer pruning method,and further compressed the model volume.Finally,this paper quantifies the trained neural network model,converts the highprecision data types in the model parameters into INT8 type data,and improves the reasoning speed of the algorithm.The algorithm is deployed on the on-board camera to construct the traffic sign recognition system for autonomous vehicles and carries out relevant experiments.The experimental results show that the prediction performance of this system is significantly improved on the task of traffic sign detection on complex road surface.