Design And Implementation Of Embedded Traffic Sign Recognition System Based On Deep Learning

Traffic-sign recognition system(TRS)is a key component of unmanned sensory system.At present,the research on TRS is to collect images of the actual driving process of vehicles in real time through cameras installed on vehicles,and then detect and identify traffic signs in the images through a variety of image processing technologies.Because real traffic scenes are more serious due to external environment impact and require high detection speed and accuracy for the application of the TRS.It is important to study TRS from theory to application.In recent years,deep learning has flourished in image processing,which provides a new idea for TRS.This paper uses deep learning methods to solve difficulties in TRS,and further complete the implementation of TRS on embedded devices.The main study content includes:This paper studies the problem of TRS in complex scenes.Most of the images in common data set are collected from a simple scene,in which traffic signs have high quality.But traffic signs in real scenes are easily exposed to external environment impacts,such as partial blockage,blurring,inclement weather,resulting in large false and missing detection.For the reason of make the model trained in this data set better adapt to complex situation in real traffic scene,this paper proposes three effective data augment methods.It is found that the model used the data augment methods in this paper has a higher detection accuracy of 89.3% and far greater than the primordial model of 77.7%,which verifies the effectiveness of image augment method done.This paper adopts deep learning to recognize traffic signs.First of all,using K-means algorithm to regroup on CCTSDB,9 priori boxes are obtained,and the YOLOv4 reference model is trained.Secondly,for the embedded implementation of TRS,a model compression method is proposed to compress the YOLOv4 reference model,reduce the size of the model and accelerate the model running speed.Using the model compression method based on channel number condense in this paper,the size of the model is cut down to 1/63 of the reference model under limit conditions,and the storage space required is reduced from 245 MB to 3.9MB.At this time,the model can still reach77.6% recognition accuracy.This paper chooses the Vision Seed as the embedded platform,deploys the traffic sign recognition model on it,and achieve TRS.First,the model conversion is carried out,the model format trained under the Darknet is converted to the format supported by Vision Seed,then the deployment of the traffic sign recognition model is completed by calling the NCAPI and loading the model parameter configuration file,and finally test the TRS.The results show the TRS run at a real-time speed of 30 FPS,and can accurately identify traffic signs in road scenes.