A unified deep learning architecture for object and motion detection in vehicle cameras

Developments in intelligent vehicles and autonomous driving technologies are crucial for ensuring safe and comfortable driving. In autonomous driving and intelligent vehicles, where driving activities are defined as partially or fully independent of the driver, various applications exist for the rapid detection and determination of object movement. Traditional algorithms used in these applications typically involve a two-step process: first detecting objects in video frames, and then tracking and analyzing their movements. This study presents a novel deep learning-based method that integrates spatial (position in the image) and temporal (change over time) information into a single architecture. The model simultaneously analyzes the objects' positions and movements to detect similar motion patterns across different object classes. These motion patterns leave traces in both spatial and temporal domains. These traces, known as Motion Profile Images, visualize the objects' movement over time and help in understanding motion patterns in complex driving environments. The thesis introduces an innovative approach that directly detects objects and their movements from a large dataset of driving videos. Specifically designed to address the challenges of dynamic driving scenarios, this method provides real-time performance. It demonstrates its effectiveness in practical applications by delivering interpretable motion detection in both spatial and temporal domains. Additionally, it confirms its reliability with high accuracy, achieving 78% average sensitivity and 3.09˚ average error on public datasets, reflecting a 22% improvement in performance.