Intelligent Recognition System Based On Raspberry Pi Embedded Traffic Sign

Traffic signs are the basis to ensure the smooth urban traffic.As an important part of intelligent transportation system,traffic sign recognition（TSR）can remind drivers or feedback to other systems through automatic recognition of traffic signs with the benefits of reduction of traffic accidents or improved traffic efficiency.It has good application prospects in the fields of traffic auxiliary safety and driverless autos with great significance for the digital city.Deep learning provides an advanced and intelligent target detection and recognition method.It has the strong abilities including feature learning,adaptability and dealing with complex tasks.Therefore,it has attracted the attention of researchers in the field of traffic target recognition.Meanwhile,low-cost small-volume and multifunctional embedded devices are widely used by vehicle terminals,traffic roadside units and so on.An embedded TSR systems can complement the existing cloud detection applications and will help to popularize their applications.However,due to the complex network structure and low efficiency,it is still facing challenges when the existing deep learning models are deployed to the general embedded systems suffering from limited computing resources.Therefore,it is very necessary to study the embedded and intelligent TSR systems.Aiming at the detection performance optimization between lightweight,detection speed and recognition accuracy of deep learning models in embedded application,based on research achievements at home and abroad,this paper proposes a high-precision and lightweight TSR algorithm SC-yolov5s-lite,and implements an embedded TSR system on Raspberry Pi platform.The main research contents and innovative contributions are as follows:（1）Firstly,aiming at the problem of the existing lightweight Yolov5 s model which may not be adapt to the small target recognition of traffic signs due to the insufficient feature extraction ability and low recognition accuracy,a TSR algorithm SC-Yolov5 s based on coordinated attention（CA）is proposed.By introducing CA mechanism,stem block structure,detection scale optimization and K-means++ clustering anchor box,the optimized network structure of new model can improve the detection accuracy obviously.（2）Further,in view of the performance balance requirements of embedded TSR application on memory,speed and accuracy,a channel-lightweight TSR algorithm SC-Yolov5s-lite is proposed on the basis of the above SC-Yolov5 s.By introducing lightweight Fire module and streamline the parameter numbers of Bottleneck residual modules,the new model is effectively compressed with improved detection speed.and more suitable for deployment on small embedded devices;（3）Finally,the Raspberry Pi embedded TSR recognition system is constructed and tested.At first,the traffic sign data set is constructed and calibrated based on the TT100 K source.Then the model training,ablation experiment and performance evaluation of the new model are completed.Finally,with the OS environment and Torch deep learning operation environment built on the Raspberry Pi 4B device,the embedded TSR system is implemented and tested gradually under the different testing conditions such as traffic scene pictures,dynamic traffic video and a real-time sensing scene of camera.The results show that the SC-Yolov5s-lite model can achieve memory occupation of 7.7m,m AP recognition accuracy of 79.7%,and Raspberry Ri embedded reasoning delay of 2.9s.In case of traffic scene pictures,the average recognition rate of 18 cases in 5 scenes can achieve 83.3%,resulting in good recognition effects of the realized TSR systems.In sum,the new SC-Yolov5s-lite model proposed in this paper has taken into account the comprehensive performances of lightweight,high detection accuracy and fast detection speed.It is more suitable for the deployment of embedded application environment with the advantages of accurate recognition,small volume,good usability,low cost and good application prospects.