This paper details the design, optimization, and evaluation of an embedded Advanced Driver Assistance System (ADAS) prototype built on a Raspberry Pi 3. It uses a custom-trained YOLOv12n object detection model. To handle the platform’s computing limits, we applied post-training INT8 quantization. This reduced the model size by over 65% and nearly tripled the inference speed, with only slight accuracy loss. The quantized model achieves about 6 frames per second in isolated inference tests. When combined with ultrasonic sensing, a lightweight lane-detection module based on classical computer vision, and GPIO-driven motor control, the complete perception and control pipeline operates at around 1 frame per second. We developed a Flask-based web interface to enable real-time visualization and telemetry during testing. Overall, the results show that optimized deep learning models can effectively support real-time perception and control on low-cost, resource-limited embedded platforms.