Research On Traffic-sign Detection Based On Convolutional Neural Network

Traffic sign detection is a vital function of Advanced Driver Assistance System,Which helps drivers obey traffic rules and ensures safe driving.Currently,most stateof-the-art traffic sign detectors are Anchor-Based,which traverse potential object locations based on anchors,and it can get a high recall rate.However,the calculation of anchor also introduces a large computational overhead.At the same time,the setting of anchor hyperparameter has a great impact on the detection performance.In order to reduce the computational overhead and excessive hyperparameter setting caused by the anchor,this thesis makes a series of improvements based on the encoder-decoder structure and the Anchor-Free detection model.The mainly 2 contributions of this thesis are as follows: Firstly,aiming at the problem that the current decoder module is simply designed,which is difficult to obtain better feature representation ability in the process of feature channel change and spatial resolution improvement,the residual enhancement branch is introduced into the decoder module,and the enhanced channel and spatial information is used as the supplement of the decoder module in the way of residual to improve the output heatmap quality of the final decoder sub-network,and in view of the disadvantage that it is difficult for a single scale feature map to fully retain the fine-grained information in space,a multi-scale feature fusion sub-network is introduced between the encoder subnetwork and the decoder sub-network,which uses the multi-scale feature information generated by the backbone network,after feature extraction and reducing the background information,the multi-scale feature information is fused as an enhancement of the decoder sub-network to further improve the detection performance of the model.Secondly,aiming at the problem that the feature extraction module in the multi-scale feature fusion sub-network introduces high parameters and computational overhead,the lightweight convolution module is used to replace the original convolution module to reduce the amount of computation,and in order to reduce the impact of the directly use of multi-level feature information with large semantic gap on the model,the deep supervision mechanism is introduced to supervise the multi-level output of the multiscale feature fusion sub-network,so as to improve the detection performance of multiscale traffic sign.The proposed model that equipped with Res Net-50 as backbone network has achieved a recall of 93.7% and an precision of 90.3% on the Tsinghua-Tencent 100 K dataset,while the parameter amount and FLOPs of the model are about 34.1M and571.2B respectively.The proposed model that equipped with Res Net-18 as backbone network has achieved a recall of 92.6% and an precision of 90.3%,while the parameter amount and FLOPs of the model are about 16.1M and 320.1B respectively.Experimental results demonstrate that our proposed algorithm has higher precision,lower computing cost,and higher comprehensive performance than the mainstream detection algorithms.