Developed an edge-accelerated computer vision analytics pipeline for automated object detection, queue telemetry, and spatial occupancy heatmaps.
1. The Engineering Challenge
Processing multi-camera 4K RTSP video streams on standard servers produced severe latency spikes (>3.5s) and frequent dropped detection frames.
2. The Architectural Solution
Engineered a TensorRT-optimized YOLOv8 pipeline running on GPU edge nodes with asynchronous video decoding and sub-30ms per-frame inference telemetry.
3. Measured Real-World Impact
- Sub-28ms per-frame inference latency enabling true real-time surveillance analytics
- 98.4% mean average precision (mAP) achieved across complex multi-object occlusion
- Streamlined edge hardware costs by 55% using custom quantization pipelines
- Interactive web dashboard delivering real-time occupancy heatmaps and automated anomaly alerts
