ADAPTIVE TRAFFIC LIGHT CONTROL SYSTEM BASED ON DEEP LEARNING

Abstract

This project is an implementation of a deep learning approach for the real-time traffic management system. Data were collected from Suryabinayak - Gatthaghar highway 1500 images which contain different classes of vehicles and mainly categorized in three classes Car, Motorbike and Truck. The custom dataset collected was used to retrain the YOLOv3 tiny model. YOLOv3 is an object detection algorithm that contains 53 convolutional neural networks, each followed by batch normalization layer and Leaky ReLU activation. The dataset was trained in Google Colab GPU. The model is tested, evaluated and saved. The trained YOLOv3 tiny model has a mAP of 82.72%, an F1-Score of 0.81, a precision of 0.80 at an IOU threshold of 0.6. The saved model was used to detect, count and manage the traffic of two lanes where fps corresponds to 2. The obtained model is Freezed, quantized and compile to make ready for deployment in Ultra96-V2 board. The obtained .Xmodel is used in inference code where input from the two USB camera for two different lanes is given and detection as well as count of the vehicles is used to control the traffic light status.