Research On Lightweight Traffic Sign Recognition Algorithm Adapted To Haze Weather

Traffic sign recognition,as a critical component of autonomous driving and assisted driving,is a key way to ensure the safety of transportation.With the development of industry,the frequent occurrence of haze weather leads to a decrease in the recognition accuracy of traffic signs that rely on computer vision.A common solution is to dehaze the image before recognition,but adding a preprocessing module is easy to lead to model enlargement,and it is not easy to deploy on vehicle devices with limited memory and computing resources.Therefore,this thesis investigates a traffic sign recognition algorithm adapted to haze weather to overcome the impact of harsh weather on image quality,and ensure that the traffic sign recognition model is lightweight,accurate and fast.The main research contents of this thesis are as follows:(1)To address the problem that low-quality images collected in haze weather lead to a decrease in object detection accuracy,an improved image dehazing algorithm based on AODNet is proposed to improve image quality.AOD-Net has a excellent dehazing speed,but its dehazing effect needs to be strengthened.To balance the relationship between the two,SE channel attention and pyramid pooling modules are introduced into the network structure and the loss function is modified to enhance the feature extraction ability of hazy images,making the details and texture features of the dehazed image more abundant.Meanwhile,depthwise separable convolution is used to replace some of the traditional convolutions in the original algorithm to maintain a faster dehazing speed.Experimental results show that the improved AOD-Net algorithm improves the dehazing effect of the image with only a small time cost.The dehazed image is well restored in color and detail texture,which is conducive to improving the accuracy of subsequent traffic sign recognition.(2)To address the problem that existing traffic sign recognition algorithms have difficulty balancing model memory,detection speed,and detection accuracy,a lightweight object detection algorithm YOLOv5s-GC based on YOLOv5 s is proposed.Firstly,Ghost module is introduced to replace C3 layer and convolution layer in YOLOv5 s algorithm to reduce model size and improve detection speed.Secondly,a coordinate attention mechanism is added to enhance the feature extraction ability and make up for the accuracy loss caused by lightweight design.Finally,the K-means++ clustering algorithm is used to obtain prior boxes suitable for the traffic sign dataset to further improve the detection accuracy.Experimental results show that the size of the improved traffic sign recognition model is only 8.9 MB,which is 37% lower than the YOLOv5 s algorithm,improves the detection accuracy by 1.2%,and accelerates the recognition speed by 24%,making it more suitable for actual driving scenarios.(3)Design a traffic sign recognition method adapted to haze weather conditions.A haze map discrimination module is introduced to distinguish the hazy images using grayscale histogram features.Hazy images are first dehazed using the improved AOD-Net dehazing algorithm and then recognized through the YOLOv5s-GC detection algorithm.The experimental results show that the accuracy of traffic sign recognition of hazy images reaches84.8%,which is 5.2% higher than that of directly identifying hazy images and effectively achieves traffic sign recognition in haze weather conditions.