Traffic Sign Detection And Recognition Based On Lightweight Multi-scale Feature Fusion And Attention Mechanism

As one of the key technologies of intelligent vehicles,advanced driving assistance systems play an important role in the safe driving of vehicles.As a key part of advanced driver assistance system,traffic sign detection and recognition can quickly and effectively perceive the traffic sign information within a certain range,and guide drivers to drive safely under the condition of abiding by traffic rules.Aiming at the problems of slow detection speed and low accuracy of existing convolutional networks,this thesis proposes a traffic sign detection and recognition method based on lightweight multi-scale feature fusion and attention mechanism.The main research work of the thesis is as follows:(1)Aiming at the problem of increasing time and space complexity while pursuing network performance in existing convolutional neural networks,this thesis proposes a traffic sign detection and recognition network based on lightweight multi-scale fusion(LMF-TSDR).Firstly,LMF-TSDR uses the lightweight feature extraction basic network to extract the basic feature information of the input image to obtain feature maps of different scales.Next,the network constructs a feature interleaving module,which converts the single-scale feature map of the specified layer into a multi-scale feature fusion map,and realizes the joint coding of highlevel semantic information and low-level semantic information.Finally,the network constructs a key point detection network to output the location information,offset information and category probability of the center point of the traffic sign,and obtain the location and category of the traffic sign through post-processing.(2)LMF-TSDR adopts compression excitation module for basic feature extraction,without considering the difference of feature map in spatial domain,resulting in a waste of computing resources.To solve this problem,the spatial domain attention mechanism is introduced on the basis of LMF-TSDR,and a traffic sign detection and recognition network combining multi-scale features and hybrid attention mechanism(MFHA-TSDR)is proposed.The hybrid attention mechanism can improve the network’s ability to extract key information about space and channels.The hybrid attention module constructs the spatial domain attention branch and the channel domain attention branch,and uses different branches to obtain the spatial and channel attention weight map.Finally,the spatial and channel attention weight map is multiplied by the input separately and then added to obtain a hybrid attention map.(3)Based on the TT100 K and GTSDB data sets,this thesis constructs a deep learning experimental platform to verify the effectiveness of the algorithm.The experimental results show that the algorithm in this thesis has a recognition accuracy of 85% for objects of different scales and most categories.Compared with Faster R-CNN and Coner Net,the algorithm in this thesis improves the performance of small object detection significantly and achieves good realtime performance.