Research On Traffic Sign Detection Method Based On CenterNe

In recent years,evolving artificial intelligence technologies have driven advances across industries,and emerging technologies related to automobiles have flourished.The emergence of the field of assisted driving has attracted many domestic and foreign research institutions to develop.Assisted driving provides drivers with information on road conditions by sensing the vehicle’s surroundings.As the main carrier of conveying the surrounding road environment information,traffic signs play an important role in ensuring traffic safety and traffic flow scheduling.Therefore,traffic sign detection technology has become one of the core components of assisted driving.The continuous development of computer vision has brought new ideas to traffic sign detection,especially the emergence of the Anchor-free detection algorithm,which reduces the arithmetic power consumption of the detection process and detects relatively fast because it does not need to design anchor frames.In this thesis,a traffic sign detection method based on improved CenterNet network is proposed to meet the needs of traffic sign detection.The main research contents are as follows:(1)Based on the full analysis of domestic and foreign traffic sign algorithms,CenterNet algorithm is selected as the object,and for the shortcomings of CenterNet algorithm applied to traffic sign detection,ResNeSt-50 is selected as the backbone network and the structure is improved.(2)To address the problem of single convolution receptive field in ResNeSt-50 structure,multi-level convolution PSConv is introduced to expand the range of receptive field of convolution check features and improve the ability of representation of traffic sign targets.To address the problem of scale variation of traffic signs,a multiscale feature module is proposed to expand the feature map size by using multiple cavity convolutions with different expansion rates,and to optimize the output feature map by combining the attention mechanism.In addition,for the problem of accuracy loss in the decoding process,a feature enhancement module is proposed to reduce the feature loss due to continuous upsampling.Finally,the size loss function of CenterNet is optimized for the disadvantage of inaccurate target size.(3)Based on the homemade dataset,the proposed traffic detection method is designed for validation experiments,and ablation experiments are taken to analyze the performance and effectiveness of the algorithm before and after the improvement.Comparative analysis of detection accuracy and speed with current mainstream algorithms is conducted to verify the effectiveness of the proposed method.(4)The improved model is deployed on the embedded platform TX2.We use ONNX as an intermediate model to convert the Pytorch model to Tensor RT,and use FP16 format to perform low-precision inference on the improved model to complete the inference acceleration of the model on the embedded platform.