Object detection explained: how AI actually ‘sees’ people and vehicles

YOLOv8 detects a person in 80 ms per image on CPU. That’s what runs inside your camera when the vendor says “AI built-in”.

A modern detection network does three things in one pass: scans the entire image, predicts where each object is, and classifies what it is. YOLO —You Only Look Once— named that idea back in 2015: treat detection as regression, not two separate tasks. The base version runs at 45 fps; the fast one, 155 fps [Redmon et al., 2016, https://arxiv.org/abs/1506.02640].

The de facto dataset for training these models is COCO: 80 categories, 330K images, 1.5 million annotated instances [COCO dataset, http://cocodataset.org/]. When a vendor says “detects people and vehicles”, they usually mean “trained on the 80 COCO categories”.

The universal accuracy metric is mAP@0.5:0.95. The YOLOv8 nano —the smallest in the family— hits 37.3 mAP on COCO val2017 at 80.4 ms on CPU or 0.99 ms on an A100 with TensorRT [Ultralytics YOLOv8 README, https://huggingface.co/Ultralytics/YOLOv8].

Your camera doesn’t run the full model on sensor silicon. What you usually get is a dedicated chip (1-2 TOPS) running a quantized version —smaller, less accurate, but enough for real-time alerts. When it logs “person detected in zone 3 at 14:32:07”, that event travels as metadata over the video stream. ONVIF Profile M defines exactly how: standardized interfaces for generic object classification plus specific metadata for vehicles, license plates, faces, and human bodies [ONVIF Profile M, https://www.onvif.org/profiles/profile-m/].

The AI in video surveillance market is projected to grow from USD 3.90B in 2024 to USD 12.46B by 2030 (21.3% CAGR) [MarketsandMarkets via PR Newswire, https://prnewswire.com/news-releases/ai-in-video-surveillance-market-worth-12-46-billion-by-2030---exclusive-report-by-marketsandmarkets-302336942.html]. The money moves because object detection already runs in production, not just in demos.

P.D. Does your camera actually “see” people and cars, or just record silhouettes nobody reviews? If the answer is “I don’t know”, the datasheet should say which detector is on board and what chip it runs on. Today’s a good day to find out.

📰 Sources:

  1. Ultralytics — YOLOv8 docs: https://docs.ultralytics.com/models/yolov8
  2. Ultralytics YOLOv8 (Hugging Face): https://huggingface.co/Ultralytics/YOLOv8
  3. COCO dataset: http://cocodataset.org/
  4. Redmon et al. — YOLO paper (2016): https://arxiv.org/abs/1506.02640
  5. ONVIF Profile M: https://www.onvif.org/profiles/profile-m/
  6. MarketsandMarkets — AI in Video Surveillance (PR Newswire): https://prnewswire.com/news-releases/ai-in-video-surveillance-market-worth-12-46-billion-by-2030---exclusive-report-by-marketsandmarkets-302336942.html

Sources

[1] Ultralytics — YOLOv8 docs (release date + architecture) — https://docs.ultralytics.com/models/yolov8 [2] Ultralytics/YOLOv8 (Hugging Face canonical README) — https://huggingface.co/Ultralytics/YOLOv8 [3] COCO — Common Objects in Context (official dataset) — http://cocodataset.org/ [4] Redmon et al. — You Only Look Once (CVPR 2016) — https://arxiv.org/abs/1506.02640 [5] ONVIF — Profile M (metadata + analytics interfaces) — https://www.onvif.org/profiles/profile-m/ [6] MarketsandMarkets via PR Newswire — AI in Video Surveillance (Dec 2024) — https://prnewswire.com/news-releases/ai-in-video-surveillance-market-worth-12-46-billion-by-2030---exclusive-report-by-marketsandmarkets-302336942.html