| Traffic sign detection and identification system is an important part of ADAS system.The system can detect the types of traffic signs in the road images and send prompt messages to drivers to reduce the occurrence of traffic accidents.At present,the relatively mature traffic sign detection system is mostly based on the complex neural network model and the cloud equipment with strong computing power.Although the system can accurately complete the detection of traffic signs,there are some problems such as network delay and unstable operation.Therefore,a traffic sign detection and recognition system based on lightweight neural network model and embedded platform is proposed to solve the problems existing in the cloud system.Due to the limited computing capacity of the embedded platform,the lightweight target detection model Yolov3-Tiny was selected and improved according to the characteristics of the traffic sign data set.Firstly,the target detection algorithms in the field of traditional machine learning and deep learning are studied.By comparing the advantages and disadvantages of the two,the advantages of the neural network target detection model are determined,and the commonly used neural network layer is introduced theoretically,laying a foundation for the subsequent research.Secondly,the collection and production of traffic sign data sets are completed.Focus structure is added to yolov3-tiny network structure to increase image information fusion.The output channel of Yolov3-tiny is added to improve the detection accuracy of small targets.In model training,teacher-student network is used for transfer learning.The resulting improved model improves the m AP by about three percent compared to yolov3-Tiny.Finally,Jetson TX2 development board was selected as the carrier and Tensorrt library was used to optimize the structure of the network model.The model parameter type was selected with low precision,which improved the detection speed of the model from 7FPS to12 FPS.Figure 44;Table 14;Reference 53... |
