Deep Learning Based Driverless Traffic Sign Detection Algorithm Research

With the extensive use of artificial intelligence technology,the major focus of research and development in the automotive industry is gradually shifting to driverless vehicles.Since traffic signs in roads are extremely important for unmanned driving as traffic guidance and regulation,the target detection technology of unmanned vehicles plays a significant role in unmanned driving.In order to optimise the performance of driverless-based traffic sign detection algorithms and improve the accuracy and speed of traffic sign detection,deep learning traffic sign detection algorithms based on driverless vehicles are investigated in depth.By introducing the two-stage target detection algorithm,the reasons for its timeconsuming detection are explained.The first-stage YOLO series target detection algorithm is then introduced,and the structure and improvements of each version of its algorithm are presented,with emphasis on the network structure of YOLOv5 s.The weak feature extraction capability of the backbone network of the YOLOv5 s algorithm is addressed,and the residual branches affect the complexity and parallelism of the model.A high-precision RC-YOLOv5s(Rep VGG-CBAM-YOLOv5s)traffic sign detection algorithm incorporating Rep VGG structure and CBAM attention mechanism is proposed,and the EIo U loss function is introduced considering the problem of vague definition of GIo U loss function aspect ratio in YOLOv5 s.The improved RC-YOLOv5 s algorithm,based on the reparameterisation function of the Rep VGG structure,improves the algorithm’s detection accuracy and real-time performance for traffic signs by fusing multi-branch image feature information in the inference stage.In order to cope with different environmental conditions during image acquisition in the unmanned process,the CCTSDB traffic sign dataset was expanded by means of data augmentation,and ablation experiments were completed on the expanded dataset to confirm the effectiveness of each improvement point.m AP values of the RC-YOLOv5 s algorithm reached 95.6%,compared to YOLOv3-tiny,YOLOv4s-mish and YOLOv5 s by 9.8%,6.2% and 4.8%,respectively,and the experimental data demonstrate the effectiveness and feasibility of the high-precision model RC-YOLOv5 s.To solve the problems of limited computing power of the embedded devices of unmanned vehicles in unmanned driving and the huge number of parameters of the YOLOv5 s backbone feature extraction network model,the Mobile Net V3 backbone feature extraction network is designed by comparing the number of parameters of the deep separable convolution with that of the standard convolution,using its deep separable convolution to achieve the goal of lightweighting the model and improving the detection speed of the network.To enhance the perceptual wildness of the network model,an RFB feature pyramidal pooling structure based on null convolution with small convolutional kernels is introduced.Considering that the large scale convolution in the CBAM attention mechanism can only obtain local information and ignore long distance dependencies,the improved algorithm embeds a CA attention mechanism incorporating location information to improve the network’s capability for the target to be detected,and continues to use the EIo U intersection and ratio loss function to improve the regression accuracy.The model is trained using a more diverse TT100 K dataset as training samples for the complex and variable traffic signs encountered in the actual operation of the unmanned vehicle.Experiments show that the improved MR-CA-YOLOv5s(Mobile Net V3-RFB-CA-YOLOv5s)model reduces the model size by 47% and increases the detection frame rate by 33% compared to the original algorithm with a 3.6% reduction in m AP,achieving an optimisation of the model parameters and real-time detection frame rate.To verify the viability and efficiency of the above two enhanced traffic sign detection algorithms on an unmanned vehicle,the algorithms are implemented on an unmanned vehicle and the detection of traffic signs is completed using Py Qt5,a GUI tool from the standard Python library,to visualise the interface design of the traffic sign detection algorithm.