Real-Time Accuracy and Effectiveness of Computer Vision and Machine Learning in Vehicle Intent Prediction at T-Junctions

Abstract

This thesis introduces a systematic approach to devising and evaluating advanced methods for predicting vehicle intent at unsignalised UK T-junctions. The primary focus of this thesis revolves around exploring the application of machine learning and computer vision techniques for the real-time prediction of vehicle intentions, emphasising increasing the prediction distance from the merge line at T-junctions to improve prediction efficacy. This thesis addresses the sparsity of publicly available data on vehicle behaviour and feature vector data at UK T-junctions by demonstrating methods to create a unique Junction Video Dataset as a foundational contribution to the field. The methodology encompasses collecting, preprocessing, and annotating video data to develop video data input for a pipeline for vehicle intent prediction. The thesis presents a comparative analysis of YOLOv5 and Faster R-CNN models, focusing on their performance in vehicle detection using the curated Junction dataset. It then introduces an innovative fine-tuning process that enhances real-time detection capabilities. This thesis uses an advanced feature extraction method to extract the stochastic nature of vehicle behaviour exhibited at T-junctions. This approach employs a sophisticated data processing and learning strategy incorporating extracted features. It continuously updates the training dataset with new feature vectors, enabling perpetual learning and the capability to make intent predictions on newly acquired data in real-time. This thesis evaluates DAISY, a real-time vehicle intent prediction model, by comparing its performance with state-of-the-art systems such as Waymo's ChauffeurNet, Tesla Autopilot, NVIDIA Drive, and Mobileye. Key metrics for comparison include accuracy, latency, robustness, and scalability. DAISY demonstrates competitive accuracy and latency, which are crucial for real-time applications. It also benefits from a modular design that enhances scalability. However, direct comparisons are challenged by systematic differences like proprietary datasets, specialised hardware, and varied algorithm complexities. This thesis revealed that integrating machine learning and computer vision techniques with high-quality data can accurately predict vehicle intent at T-junctions. Such an approach has the potential to serve as a crucial element within a safety model, functioning as an early warning system or activating driver assistance features.