Development of Deep Learning Frameworks for Object Detection and Flow Prediction in Traffic System

Abstract

Road traffic object detection and prediction play a crucial role in Intelligent newlineTransportation System (ITS) for various approaches, like traffic management, driver newlineassistance systems, in addition to autonomous vehicles. By using computer vision techniques, newlineobject detection and traffic prediction algorithms can identify and track objects newlineand predict in real-time, enabling ITS systems to make informed decisions and take newlineappropriate actions. Object detection and prediction has various applications specially newlinein autonomous driving, traffic surveillance, traffic management and safety, advanced newlinedriver assistance systems (ADAS). Object detection and prediction is done through still newlineimages, frames captured from webcam and surveillance cameras. However, an image newlineobtained by the image acquisition system is overwhelmed by uncontrolled environmental newlinecomponents including traffic congestion, flow and poor weather conditions. These newlineadverse conditions result in a significant decrease in image or frame quality hindering newlinedetection system from discerning the target surface s reflection. This prevents the detection newlinesystem from fully confirming the detection and prediction of an traffic management newlinesystem. newlineTo address the problem of traffic object detection and prediction, techniques newlinesuch as TensorRT and Triton Inference techniques can be used to improve the visual newlineexperience while making outdoor visual systems more reliable and durable, resulting in newlinemore accurate object detection and prediction. Studies on traffic object detection and newlinepredictions use the conventional methods and deep learning models are few. However, newlineobject detection and prediction with conventional methods and deep learning models newlinehave issues of less discrimination and artefacts of the traffic prediction, which lead to newlinepoor performance in vehicle object detection, traffic flow, congestion, and travel time newlineestimation. newlineHence, the first work in this thesis investigates the accuracy of object detection newlineand proposes an architecture for object detection.