Traffic Detection for Advanced Driver Assistance System

Abstract

The Advanced Driver Assistance System (ADAS) is not a new phenomenon. To minimize road accidents and other related issues, the current vehicles can be improved for a better driving experience through an automated system that assists the driver. Some of the basic elements that such ADAS systems utilize include, but are not limited to, sensing the environment, traffic signs, pedestrians, and other vehicles. The need for traffic to be detected and recognized up to a certain degree of accuracy arises due to our objective i.e., to ensure that the car and the passengers in it are safe. Traditional Image Processing techniques have previously been used which are way slower. Recently, CNNs have been deployed heavily in Traffic detection and identification. However, CNNs do require a huge number of input images to work efficiently, and no such traffic recognition dataset exists in Pakistan. In this research, we deployed a YOLOv7 based architecture trained on a self-collected and manually annotated Pakistani Traffic Type and Sign Recognition Dataset (PTSD) to detect and classify the types of traffic. The Deep Learning model was trained and tested to produce a mean average precision (mAP) of 87.20%. These results are state-of-the-art and strong enough for implementation as real-world models. The model was further tuned to help improve the model’s working, and then tested in real-world scenarios. The final model was used to develop an ADAS Unit— which works on a priority-based decision system, providing specified instructions for the detected conditions.