Research And Application Of Driverless Oriented Traffic Signal Detection And Recognition Method

With the rapid development of computer technology and communication technology,driverless driving is no longer a mere guess,and more and more intelligent technologies of the Internet of Things are being used in cars.As a key technology of unmanned driving,the detection and recognition of traffic signals plays a key role in the follow-up behavior of vehicles.In unmanned driving,the front traffic signal needs to be quickly and accurately obtained and transmitted to the decision-making layer.The decision-making layer analyzes and judges and transmits it to the control layer,and the control layer controls the behavior of the car.With the increase in the scale of deep learning networks,the number of parameters becomes very large,which requires a large amount of memory space and calculations.Therefore,it is necessary to study a lightweight traffic signal detection and recognition model that can run smoothly.When a car is running,it needs to detect and recognize traffic signals at a long distance.The size of traffic signals at long distances is small,so it is necessary to study traffic signal detection and recognition models for small targets in complex environments.The detection and recognition of traffic signals need to meet the requirements of accuracy and speed.The main content of this thesis includes the following aspects.(1)A lightweight traffic signal detection and recognition model is proposed.This model uses a lightweight network to Feature extraction of traffic signals reduces the demand for computing resources by reducing model parameters.At the same time,an improved lightweight attention mechanism is introduced to improve the detection accuracy of traffic signals.The model was tested on the data set,and the detection speed of the proposed model was increased by 26% compared with the original model,and the size of the obtained model was reduced by half compared with the original model,which confirmed the usability of the model.(2)Propose a small target traffic signal detection and recognition model.The model draws on FPN feature gold characters Tower and PANet’s multi-scale prediction ideas,aiming at the feature of small targets,use up-sampling and feature map fusion splicing to construct a multi-scale target detection network suitable for small targets.At the same time,the clustering algorithm is introduced to adjust and determine the appropriate prediction frame.The model is tested on the data set,and the proposed model has improved the detection accuracy by ten percentage points,and the detection accuracy has been greatly improved,which confirms the usability of the model.(3)Design and implementation of traffic signal detection and recognition system.The system is based on the design of unmanned driving perception platform and realization.As a very important traffic signal detection and recognition module of the driverless perception platform,this system mainly realizes the two functions of local offline detection and online traffic signal detection and recognition.Local off-line detection mainly uses local video files to be transmitted to the model proposed in this paper for detection and recognition,and the results are displayed.On-line traffic signal detection and recognition mainly uses real-time images taken by the camera and transmitted to the model proposed in this thesis for detection and recognition,and the results are displayed.The test results show that the proposed model can be deployed on the driverless perception platform,further confirming the usability of the model.