Research On Urban Road Traffic Signal Recognition Based On Deep Learning

In recent years,with the rapid innovation of science and technology and artificial intelligence,autonomous and assisted driving technologies based on deep learning and computer vision are gradually replacing traditional driving methods to improve road traffic safety.And traffic signals are one of the important elements that make up road traffic safety.Efficient and accurate identification of the status and category of traffic signals can help intelligent cars obtain traffic intersection information in advance,which is conducive to the intelligent system to make correct driving decisions and avoid major traffic accidents.Therefore,in this article,an improved algorithm based on YOLOv5 is proposed for the detection,recognition and application of traffic signal images,taking domestic urban road traffic signals as the research object.The main research contents include the following aspects:(1)The establishment of domestic traffic signal datasets.Through the analysis of the currently open source traffic signal datasets,it is found that most of them have fewer categories,and are only classified by color,which cannot provide more accurate traffic information for intelligent vehicles.In this paper,a real domestic urban road traffic scene dataset TLD covering several cities under different angles,climate and lighting conditions is established by various means such as video sampling.The dataset is meticulously divided into 9 categories according to the color and direction of traffic signals,with a total of 24,315 labeled samples obtained.(2)Research on traffic signal detection and recognition algorithm based on improved YOLOv5.Firstly,by comparing the MobileNet series of lightweight networks containing deeply separable convolutions,the MobileNetv2 network with the best detection effect is synthetically selected to replace the original backbone feature extraction network,which is utilized to achieve the lightweight of the model.Secondly,a four-scale detection mechanism is used to better cover traffic signals of different sizes and locations,and to improve the sensitivity of the algorithm to small target features.Finally,an attention mechanism is introduced to enhance the capability of the network to capture traffic signal features from both channel and space dimensions,thus improving the detection accuracy of the algorithm.The experimental results show that the improved algorithm in this paper reduces the model size by 25.4% compared with the original YOLOv5 algorithm,improves the detection accuracy by 2.81%,and achieves a detection speed of 30.6 frames per second.(3)Design and implementation of the traffic signal detection and recognition system.Based on the improved YOLOv5 algorithm,this paper designs and implements a traffic signal detection and recognition system.The system utilizes the interface development tool PyQt5 to design a visualization module for the acquisition,input,detection to output process of urban road traffic signals.Through experimental,it is proved that the system has good practical application and can be applied to traffic signal recognition under complex traffic scenarios on urban roads.