SPEEDY: Small Prototyping for Number Plate Detection and Speed Estimation using Edge AI

• Main Authors: Singha, Shivam, Saha, Debajit, Sarma, Pallab, Singh, Pranav Kumar, Ghosh, Uttam, Brahma, Maharaj
• Format: Conference Proceeding
• Language: English
• Subjects: Computational modeling; Edge AI; Estimation; Feature extraction; Licenses; Machine Learning; Prototypes; Raspberry Pi; Speed Estimation; Tensorflow Lite; Webcams; YOLO
• ISSN: 2771-5663
• DOI: 10.1109/CloudNet59005.2023.10490026

This paper presents a novel system that harnesses the capabilities of Edge Artificial intelligence (AI) to revolutionize number plate detection and speed estimation by significantly reducing costs associated with existing solutions. The system employs a low-cost Raspberry Pi device to capture and process video feeds obtained from a low-cost USB webcam or a MIPI-based camera module. Leveraging a quantized YOLO model, the system accurately detects vehicles on the road and utilizes a rectification layer to correct license plate images. Using an OCR model, the system extracts license numbers and estimates vehicle speeds through distance calculation methods. The system incorporates a pyramidal search technique that scales features instead of the entire image to enhance accuracy, resulting in improved license plate detection. Furthermore, we conducted tests to estimate vehicle speed based on changes in centroid position between frames, revealing its potential for accurate estimations, particularly in sparse traffic conditions. Our work demonstrates the effectiveness and feasibility of such a low-cost system in real-world scenarios. It shows the potential of such small prototypes in various applications such as traffic management, incident detection, and emergency services coordination that demand real-time analysis and response. The paper also discusses the system's limitations and challenges while offering suggestions for future work, including potential improvements and extensions.