Design Of Embedded Traffic Sign Recognizer

Traffic sign is a wind vane provided by road traffic to drivers,ensuring smooth running of urban traffic.With the combination of artificial intelligence technology and ITS,the detection and recognition of traffic signs has gradually become a hot issue in the field of computer vision.Its main application is in specific traffic scenarios such as intelligent transportation,autonomous driving,and assisted driving.Remind the driver by identifying the traffic signs in the video stream,and reserve enough reaction time to control the car's movement.In recent years,GPU computing-based deep learning target detection models have greatly improved the accuracy and speed of detection.More and more researchers have shifted the detection and recognition of traffic signs to lightweight and compact embedded systems.Designed to design a traffic sign recognition device that is separate from the PC and the cloud,and solves the problems caused by network delay and excessive hardware volume.This paper studies and designs an embedded traffic sign detection and recognition system.The main research contents are as follows:1.This paper analyzes the principles and characteristics of image processing algorithms and target detection algorithms commonly used in the field of target detection,highlighting the advantages of deep learning algorithms such as high detection accuracy,fast speed,and strong universality.;This article independently established a traffic sign data set,containing more than 15,000 real traffic scenes and annotation information for each image.2.In the implementation plan of the system,this paper selects the YOLOv3 target detection model based on the direct regression method,taking into account the balance between recognition accuracy and recognition speed.In order to improve the problem of low recall rate and high miss rate of small targets for the model,this paper improves the YOLOv3 model: the Gaussian clustering algorithm is used instead of the k_means clustering algorithm to select Anchor Boxes,and the IOU parameters of the pre-selection box are increased;1 * 1 convolution kernel expands the model depth and enhances the model's non-linear expression ability;cooperates with SSD multi-scale expansion detection feature spectrum,enhances the model's ability to extract detailed features.The experimental results show that the improved YOLOv3 model has a detection accuracy mAP of 0.9057 for traffic signs,which is 3.8% higher than the unmodified model.3.In order to adapt to the transplantation of embedded platforms,this paper uses the method of deep separable convolution to lighten the model based on the improved model,and accelerates the actual detection speed from less than 1FPS to the premise of not reducing the detection accuracy 5FPS.This article further improves the detection speed of the model on the hardware platform through the GStreamer video codec component and TensorRT inference optimizer,from 5FPS to 15 FPS,which can smoothly handle various actual traffic conditions.In this paper,after implementing the traffic sign detection model,we consider the complexity of the actual traffic conditions.Field model tests were performed in four cases where multiple traffic signs existed,the highway view of the car was high,traffic conditions were complex,and traffic signs were blocked.It is found that in the poor visual field environment,small targets are missed and the marker frame is repeated a little.However,in the case of good visual field conditions,the model can basically detect the target,and it has a certain robustness to the occluded target.