Self-Driving Car Using Image Processing and Real-time Perception, Lane Keeping, and Dash Board Integration

Abstract

Self-Driving Car demonstrates an integrated prototype for basic autonomous driving functions by combining real-time image-processing perception with vehicle state data obtained from the OBD-II interface. The system uses a single forward-facing camera to perform multi-task visual perception: lane detection (perspective transform, edge detection, sliding-window polynomial fitting with temporal smoothing), object and traffic-sign detection (single YOLOv8 model), and optional OCR for numeric sign readout. Visual outputs are fused with vehicle telemetry (speed, RPM, fuel) obtained via OBD-II to produce a unified HUD overlay suitable for windshield projection or driver assist. The software architecture is multi-threaded: a low-latency video pipeline performs image preprocessing and lane fitting, an asynchronous detector throttles heavier neural-network inference to maintain real-time framerates, and a background OBD reader supplies vehicle state. We evaluate the prototype qualitatively on recorded and live road scenes and report typical CPU/GPU framerates and detection latency for a YOLOv8-nano configuration. Results show robust lane tracking on well-lit paved roads, reliable detection of common road agents (vehicles, pedestrians) and traffic lights/signs, and seamless HUD updates synchronized with simulated OBD telemetry. Limitations include sensitivity to adverse illumination, missing specialized sign classes in COCO-trained detectors, and the need for camera calibration for metric lateral positioning. Future work includes training specialized sign/tree detectors, integrating IMU/GNSS for sensor fusion, adding closed-loop lateral control, and safety validation on instrumented test vehicles.