CenterNet-based Traffic Sign Detection And Application

Object detection plays an important role in computer vision tasks.Its main function is to detect a certain category of instance object in digital image and solve the problem of "where" and "what" of the object.Traffic sign detection can be regarded as a typical problem about object detection.As an important part of vehicle intelligent driving assistance system,traffic sign detection can effectively help drivers to make scientific decisions and solve common traffic problems in actual travel.This thesis studies traffic sign detection based on deep learning,and realizes the construction of embedded system with the combined research results for the needs of engineering application.1.CenterNet-based traffic sign detection by incorporating attention mechanism and multi-scale feature prediction.By introducing keypoint-based object detection,the baseline model for traffic sign detection based on CenterNet is constructed and implemented.Furtherly,by incorporating convolutional block attention module(CBAM)and multi-scale feature prediction(FPN)strategies into the Center Net mainframe,an enhanced traffic sign detection model based on CenterNet is constructed.Compared with the CenterNet-based baseline model,the detection precision of the enhanced CenterNet-based model is improved with the slightly increased training time.With the given information characteristics of domestic traffic signs,some sample images from both datasets(TT100K and CCTSDB)are selected for sample expanding via data augmentation strategy.The traffic detection experiments performed on the expanded dataset demonstrate the effectiveness of the proposed algorithm.2.Lightweight and embedded deployment of traffic sign detection model based on CenterNet.In order to meet the need of end-to-end detection,MobileNet V2-based backbone network is used instead of Res Net-based backbone network,and therefore the lightweight of the enhanced CenterNet-based traffic sign model is realized.Then,the lightweight CenterNet-based model is deployed in Jeston TX2 embedded device to meet the requirements of engineering application,and the construction of lightweight traffic sign detection system is realized.The experimental results show that the training time of the lightweight model is greatly shortened and the detection speed is improved without obvious loss of detection accuracy.