Traffic Sign Detection Algorithm Based On Convolutional Neural Network

In recent years,China’s road traffic has developed rapidly and traffic conditions have become more complex.Automated driving technology based on target detection has gradually become the focus of research.As one of the most important components in the field of autonomous driving,traffic sign detection has received wide attention from all walks of life.The target detection algorithm based on convolutional neural network is one of the most effective solution to traffic sign detection problems.Traffic signs have high requirements for detection accuracy and detection speed.The lightweight network represented by Tiny-YOLOv3 satisfies the requirements of detection speed,but the detection accuracy is generally low.The detection accuracy of large networks represented by YOLOv3 It has achieved good results but it is difficult to meet the test speed requirements.In order to solve the above problems,this paper proposes optimization schemes for Tiny-YOLOv3 light network and YOLOv3 large network:Aiming at the problem of low detection accuracy in traffic sign detection for lightweight target detection networks such as Tiny-YOLOv3,this paper proposes a convolution module based on Depthwise Separable Convolution,which effectively reduces the amount of computation required for convolution operations.Therefore,under the same calculation conditions,more levels of feature extraction networks are constructed to extract deeper image features.Meanwhile,this paper sets up more levels of feature fusion network to effectively improve the detection accuracy of small and medium-sized traffic signs;using the h-swish activation function instead of the traditional Re LU activation function effectively avoids the occurrence of gradient disappearance.The experimental results show that the Tiny-YOLOv3 network based on deep separable convolution has obtained the mean average precision(m AP)of97.54% and the detection speed of 201.5fps,which is 14.01% higher than the m AP of Tiny-YOLOv3,effectively improving the accuracy of traffic sign detection.In order to solve the problem that large-scale networks such as YOLOv3 have large computation and slow detection speed in traffic sign detection,this paper proposes a Inverted Residuals Depthwise Separable Convolution module to replace the residual module in YOLOv3 network,which is effective reduced the amount of computation of the network.In order to avoid the gradient disappearance,the feature processing module in YOLOv3 is optimized;in order to further improve the detection accuracy of small and medium-sized traffic signs,this paper adds a feature fusion layer based on 4 times downsampling and resets the feature layer on each scale feature.The number of network layers.The experimental data set of the YOLOv3 network based on the Inverted Residuals Depthwise Separable Convolution structure proposed in this paper has obtained an m AP of 98.95 and a detection speed of 87.9 fps,54.1 fps higher than the detection speed of YOLOv3,which effectively improves the detection speed of traffic signs.