ROCmVision: AI Traffic Violation Detection

ROCmVision is an intelligent traffic violation detection system built using custom-trained YOLO models, computer vision, and AMD ROCm acceleration on MI300X GPUs. The project focuses on detecting real-world traffic violations such as helmet violations, triple riding, and vehicle identification using automatic number plate recognition (ANPR). The system combines multiple AI models into a single pipeline. A custom helmet detection YOLO model identifies riders wearing or not wearing helmets, while another YOLO model detects motorcycles and riders. A dedicated ANPR detection model identifies license plate regions for further processing and future OCR integration. The project was initially developed and tested in Google Colab and later optimized to run on AMD Developer Cloud using ROCm-enabled PyTorch. By adapting the codebase for ROCm compatibility, the same inference pipeline successfully ran on AMD Instinct MI300X GPUs with high-speed inference performance. The system processes both images and video streams and generates structured outputs containing: - Number of riders - Helmet violations - Triple riding detection - Motorcycle grouping - License plate localization Performance benchmarking showed fast inference speeds suitable for real-time deployment scenarios. The modular architecture allows future extensions such as OCR integration, real-time surveillance pipelines, smart city analytics, and AI-assisted traffic enforcement systems. Technologies used include: - YOLOv8 - PyTorch - OpenCV - AMD ROCm - AMD Developer Cloud - Python This project demonstrates how AMD ROCm infrastructure can accelerate practical AI computer vision applications while maintaining compatibility with existing PyTorch-based workflows.