Traffic Lights Detection Method Based on the Improved YOLOv5 Network

Xumeiqi Chen

20222022 IEEE 4th International Conference on Civil Aviation Safety and Information Technology (ICCASIT)16 citationsDOI

Abstract

Helping people with limited vision to identify traffic signs is important and challenging work. In recent year, many methods have been proposed to detect traffic signs, but all of them have low accuracy rate, long detection time and single detection object. The most advanced one is Yolo series, and this paper proposes a method to improve the detection accuracy while shorten the detection time based on Yolov5. And this study used methods like resizing and normalization, etc. to process the LISA dataset collected from Kaggle and divide them into 8:2. Then, this work establishes a loop that updates the parameters using a gradient of loss values. Under this new model the mAP is 0.988 within 2.4ms inference/1.0ms NMS per image, which shows that this method is more suitable and more robust and more practical than Yolov5s and can be used for real-time traffic light detection.

Topics & Concepts

Computer scienceNormalization (sociology)Object detectionArtificial intelligenceInferenceProcess (computing)Computer visionData miningReal-time computingPattern recognition (psychology)SociologyAnthropologyOperating systemAdvanced Neural Network ApplicationsIndustrial Vision Systems and Defect DetectionVideo Surveillance and Tracking Methods