Fast detection method of traffic signs in multi- scenarios of deep learning
DENG Leping;LI Wei;Jihe (Zhejiang) Technology Co., Ltd.;
In order to solve the problems of error detection, missing detection and slow speed in the detection of traffic signs in various complex scenarios, this paper puts forward a traffic sign detection method based on convolutional neural network. Firstly, it uses a single-stage target detection network framework, and a combination of hollow convolution kernel and 1×1 small convolution kernel to construct a feature extraction network to obtain feature maps of different sizes. Then, the feature enhancement is carried out by multi-layer feature pyramid, and target detection results of different sizes are output through multiple detection terminals. In order to improve the model detection accuracy after training, the image pre-processing-enhancement module is constructed to normalize the input image, reduce noise and filter on the input image, and meanwhile, K-means++algorithm clustering is used to obtain the initial candidate box that best fits the actual target size. The experiment results show that the method introduced in this paper can detect traffic sign targets quickly and accurately in different natural scenes, the average accuracy of single-category targets is 94.4%, the average accuracy mean is 93.45%, and the detection speed in test environment is 31 FPS/m·s~(-1), which can realize real-time detection of traffic signs.