IMPLEMENTATION OF MACHINE LEARNING ALGORITHMS FOR IMAGE CLASSIFICATION ON EMBEDDED SYSTEMS

This project presents the implementation of machine learning algorithms for image classification on embedded systems with an integrated vehicle counter. The increasing demand for intelligent systems capable of real-time decision-making has accelerated the integration of machine learning models into embedded platforms. The system was developed on a Raspberry Pi 5 (8GB RAM) using a Logitech 720p USB webcam as the input device and a 7-inch HDMI display for real-time visualization. A pre-trained YOLOv8 model based on the COCO dataset was deployed to detect and classify vehicles such as cars, buses, and motorcycles. A line-crossing algorithm was incorporated to provide accurate vehicle counting. The project achieved a classification accuracy of approximately 86-87% across selected classes with real-time performance averaging 17 frames per second (FPS). Latency analysis showed that inference remained the most significant contributor to processing time, while optimization through resizing improved overall throughput. The prototype was tested under varying conditions and demonstrated robustness in detecting and counting vehicles with minimal errors.