Optimization Of Multi-target Traffic Sign Detection Model For Faster R-CNN

At present,with the rapid development of intelligent transportation system,the intelligent driving assistance system plays an important role,and the traffic sign detection is one of the key technologies,which can prompt the driving of the driver.Due to the difficulty of extracting features of traffic signs and the high complexity of feature extraction network,the detection effect of traffic signs is not good.Therefore,improving the representation ability of the network has become the decisive factor for the performance of traffic sign detection.Traffic signs belong to small targets and have the characteristics of multi-scale,which increases the difficulty of traffic signs detection.How to improve the accuracy of small targets detection has become an urgent problem to be solved at present.The traffic sign detection model based on Faster R-CNN have shown great advantages.Therefore,in response to the above-mentioned problems,this dissertation conducts research from the perspectives of enhancing the feature extraction performance of the traffic signs,reducing the network complexity,and constructing end-to-end multi-scale detection model.The advantages of the model proposed in this dissertation are analyzed through experiments.The concrete research content are summarized as follows:(1)Aiming at the problem of inadequate representability of traffic signs in natural scenes and high complexity of feature extraction network,a traffic sign detection model Des FR-CNN with optimized feature extraction is proposed.In this model,the original feature extraction network is replaced by the lightweight densenet,and the cross reuse of features is enhanced by using cross-layer dense connectivity pattern,which greatly improves the utilization of features.The feature dimension and learning parameters are compressed through the transition layer,thereby reducing the network complexity.Compared with the original algorithm,the m AP of the proposed model on the GTSDB dataset is increased by 5.8% and the number of parameters is reduced by 94.2%.Experimental results show that this model can enhance the representation ability of traffic signs and reduce the complexity of network.(2)The traffic sign is a small target,which increases the difficulty of traffic sign detection.In order to solve this problem,a channel attention network ECANet is proposed.The network adaptively learns the correlation between feature channels through weighted fusion,and the local cross-channel interaction without dimension-reduction was realized by using Conv1 d,so as to pay more attention to the feature information of small targets.The network was evaluated on the CTSD dataset by ablation experiments,and the m AP reached 93.1%.Experimental results show that this network can enhance the feature extraction ability of small targets.(3)To solve the problem of low accuracy in detection of multi-scale traffic signs,a multi-scale feature fusion model D-FPN is proposed,which combines feature pyramid and dilated convolution.In this model,deep strong semantic information and low-level high-resolution information are fused through feature pyramid to generate high-resolution multi-scale semantic features.The dilated convolution is used to learn multi-scale context information to reduce the difference between features,so as to improve the detection accuracy of traffic signs.Finally,the m AP on GTSDB and CTSD datasets reached 96.7% and 94.0%,respectively.The experimental results show that this model can improve the detection performance of multi-scale small target traffic signs.