Deep Learning based Object Detection for Vehicles in Local Urban Traffic

Abstract

The Project focuses on developing a model and a system to detect local urban traffic using a custom locally captured dataset. It captures video feed from a vehicular mounted Camera on a rotatable stabilized platform, which streams the captured video over a 4G network using GSM protocols to the Command-and-controlled center where the stream is processed and objects are detected using DL techniques. Which are then viewed using websites, monitors, and VR Googles, The Platform is controlled with a wireless joystick available in the centers. The knowledge parities of developed and developing countries have widened the gap in all areas exclusively technology. The deep learning-based detection of objects of different classes has powered a pathway for working in the area of Deep Learning-based Object Detection for Vehicles in Local Urban Traffic. The given project is the continuation of an attempt with the same theme based on contextual gaps. In recent years, researchers have proposed multiple object detection techniques for vehicle detection on datasets such as Kitti, Berkeley deep drive, oxford robot car dataset, nuScenes, Daimler urban segmentation (DUS), and cityscapes, however, accurate detection of objects in self-driving vehicles remains a challenge in a real-time environment with traffic on the roads. Although they had advanced technology still, we were able to achieve high accuracy by removing the misclassification problem. A misclassification can cost a human life. Therefore, we have worked on the detection of objects such as pedestrians, cars and bicycles etc. for autonomous vehicles with high accuracy. Our Proposed method is YOLOv8 which will be able to detect objects with high accuracy and less inference time. We have used the CarlD dataset for our local area.