Applications of artificial intelligence in real-time video analytics

Illegal parking can be a ubiquitous concern faced by urban cities, posing potential traffic impediments and safety risks to other road users. Despite having surveillance systems deployed to monitor traffic offences, the videos recorded are often stored only for post-event forensics. Manually inspecting the videos often involves repetitive human labour, which is tedious and prone to errors. In this project, a fully automated pipeline to perform end-to-end illegal parking detection with minimal or no human-in-the-loop was proposed. The pipeline first consists of vehicle detection using a deep learning based object detection algorithm, You Only Look Once Version 3 (YOLOv3), to detect vehicles. Next, movement tracking using template matching and Intersection over Union (IoU) are performed to track the time since the violating vehicle has remained stationary. The last step is to extract the license plate, using OpenALPR, of the violating vehicle which has remained stationary for a defined period. With the fully automated pipeline in place, the dataset can be intelligently leveraged and the analysis can be automated in real-time. Empirical results show high accuracy of vehicle detection and movement tracking module with the license plate detection module achieving a decent performance. However, improvements can be made by retraining its underlying license plate detection and Optical Character Recognition (OCR) engine with the dataset from the location which the system is to be implemented on.Bachelor of Engineering (Computer Science