Edge AI Computer Vision Prototype

I used GStreamer to separate capture, preprocessing, inference, post-processing, and output. YOLO handled object detection for vehicles, pedestrians, and environmental elements. Filtering and non-maximum suppression converted raw model output into the detections the rest of the pipeline could consume.

The staged design made the prototype easier to reason about. Video handling, model inference, and visualisation have different performance and debugging needs. Keeping them modular meant one stage could change without rewriting the whole flow.

The first output was a processed video stream with bounding boxes. The second was structured detection metadata for every frame. That metadata drove simplified ambient-light signals representing nearby objects or hazards.

The light feedback was intentionally a prototype interaction, not a production driver-assistance claim. It demonstrated how a local perception pipeline could feed a lower-attention interface while preserving a visual stream for debugging and review.

Running the pipeline on embedded ARM and x86 hardware made latency, model size, video throughput, memory use, and stage boundaries visible. There was no cloud service to absorb inefficient preprocessing or hide a stalled frame path.

The project taught me to treat computer vision as a complete streaming system. Model accuracy matters, but so do frame movement, conversion cost, post-processing, output contracts, and the way downstream components consume detections.