Research On Traffic Light Detection And Recognition Based On Vehicle Multidimensional Input Information

Traffic signal detection and recognition is one of the important functions of automatic driving environment perception.In the process of automatic driving,the vehicle needs to independently judge the indication status of the traffic lights according to the various sensors it carries,so as to provide input information for the decision-making and control of automatic driving.In the actual recognition scene,the traffic lights are usually in a complex and changeable environment.The existing research methods are difficult to give consideration to both accuracy and real-time,and cannot adapt to the actual engineering needs.Therefore,based on the actual project,this paper has carried out in-depth research on traffic signal detection and recognition,and proposed a signal detection and recognition method based on multi-dimensional input information.The main research work is divided into the following two parts:(1)An improved YOLOv5 m traffic signal detection algorithm based on visual input information(YOLOv5m std)is proposed.In order to enhance the feature extraction ability of the model for small target of signal light,the detection performance of different size feature map input is investigated experimentally,and the input scale of the model is increased in the preprocessing stage;The ACBlock convolution module and the point-surface separation convolution based on CPConv are introduced to enhance the ability of boundary feature extraction and reduce the calculation amount of the model;Co Atten coordinate attention mechanism is applied to effectively capture the relationship between images across channels;In order to make the target prediction frame closer to the real frame,ED is designed-δ UIOU positioning loss function improves the regression accuracy of model positioning to a certain extent.(2)GPS/INS integrated navigation system is selected as the system scheme of high-precision positioning data information input in this paper.The method of cubic spline interpolation is used to insert possible path points between two sampling points to obtain a complete lane centerline path,so as to build a road-level high-precision map.The concept of distance threshold is put forward,and the optimal distance threshold for the signal lamp system is explored experimentally.Through the real vehicle road experiment,the influence of the threshold setting of the opening distance of the signal light recognition system on the signal light recognition effect in the day and night environment is discussed.The test results show that this method can save a lot of computing resources and effectively improve the recognition accuracy of signal lamps.By integrating the data sets of road traffic lights in various complex environments,this paper conducts a comparative experiment with the traditional YOLOv5 m algorithm,and analyzes the relevant evaluation indicators of detection speed and detection accuracy in detail.The experimental results show that the detection speed only differs by 2 FPS,the improved algorithm has improved by 5.3percentage points compared with the original model,and the average recognition accuracy has reached 91.7%.When the positioning input information is added,the recognition accuracy has increased by 5.12 and 9.06 percentage points in the daytime and at night,respectively,which proves the superiority of the traffic signal detection and recognition method proposed in this paper.