Traffic Lights Detection And Recognition Based On Deep Learning

In the current era of rapid economic development,in order to better promote the improvement of people’s living standards and the development of science and technology,the country has issued a number of policies,including numerous regulations closely related to road traffic safety.As we all know,nowadays,the pressure on road traffic is becoming increasingly severe,and people’s pursuit of intelligent technology is also unprecedented.In recent years,the rise of deep learning has led to the rapid development of algorithms,applications,and technologies related to it.Among them,the development work in the field of intelligent driving has always been carrying out innovation and breakthroughs,and is committed to providing ideas for research in the field of intelligent driving.This thesis takes the main representative of road traffic signals-traffic lights as the research object,solves the problems existing in the current field of traffic lights detection and recognition,proposes a traffic lights detection and recognition method with high detection accuracy and meets certain real-time requirements,which is based on LISA traffic lights dataset and self-made traffic light dataset.The target detection algorithm YOLOv5 is used to study the detection and recognition methods of traffic lights.(1)a feature extraction network integrating CBAM attention mechanism is proposed.Usually,in the feature extraction network section,it is expected to capture more data features of images,so the design of the network is of great importance in improving feature extraction capabilities.In this section,based on the original backbone feature extraction network structure,a lightweight CBAM attention mechanism is integrated to further enhance the network’s ability to express traffic lights features,providing rich previous information for subsequent networks,in order to achieve high-precision detection of traffic lights without increasing network complexity.(2)a feature enhancing CSPT module embedded in Transformer structure is proposed.Due to the fact that the correlation between contexts is often different when networks process feature information,in order to better utilize the extracted features and avoid being limited by local features,it is expected that the network not only focuses on global features,but also focuses on important information.A Transformer structure with self attention mechanism is embedded in the feature fusion network,forming a CSPT module together with the original CSP network,Utilizing a model incorporating a Transformer structure to improve training speed and complete the mining of deep information,providing a basis for subsequent predictive classification.(3)a new small target detection scale is proposed.When conducting traffic signal detection tasks,due to the diversity of training sample features and categories,a multi-scale detection method was proposed to better detect small target traffic lights.The output features of the fourth layer of the network were extracted by combining the main features.After fusion with the CSPT module,the new detection scale was used to improve the detection accuracy of small target traffic lights.The experimental results show that the traffic lights detection and recognition method proposed in this article has significantly improved detection precision,recall rate and average precision(m AP),also can meet certain real-time requirements,which has important application value.