Dashboard — CCTV Intelligence
Automated person, vehicle, crowd, face and material analysis across four
camera locations at the Khavda solar plant.
YOLO26x ByteTrack
InsightFace Buffalo_L
YOLO26m fine-tuned — materials
OpenCV CUDA
Generated 29 Jul 2026, 15:20 · Unada Technologies · Khavda Solar Plant
How to read this dashboard
Each card is one processed video. Counts are unique tracks (not per-frame detections),
and material counts are per-frame observations reported as median/peak — never summed.
Full caveats are at the bottom of the page.
6 Videos analysed
42 Person tracks
15 Vehicle tracks
6 Crowd events
11 Face crops
118 Material frames
11.71 Avg FPS
Material Central Store_10.196.104.44_Recording 2026-07-17 182354
Material Central Store_Recording 2026-07-17 180127_10.196.104.47
Material Central Store_Video Pan_10.196.104.46_17.07.26
Detection PSS1 Gate - 10.208.136.11 - 2026-05-24
Detection Crowd ×6 Labour Colony_10.208.128.68_Recording 2026-07-17 185852
Detection vlc-record-2026-04-24-12h26m00s-rtsp___10.208.128.76_pro
How to read these results
Detection & tracking — YOLO26x detects people and vehicles; ByteTrack assigns each a persistent ID, so counts are unique tracks , not per-frame detections.Faces — quality-ranked best crops per tracked person (frontality × sharpness × size). Identities are Unknown : no face gallery is enrolled, so this is detection, not recognition of named individuals.Crowd events — a rules-based geometric check, not a model: 3+ people grouped within proximity for 5+ continuous seconds.Materials — Central Store now uses a custom YOLO26m fine-tuned on 93 hand-labelled frames from these cameras (Steel_Bundle, Wooden_Pallet, Scrap, Long_rebars_wire, Truck, steel_coll and rarer classes). Vehicles come from the stock COCO detector, which is far stronger there than 92 hand-drawn boxes.Material counts — per-frame observations of the same physical stacks. They are reported as median and peak and are never summed across frames , which would count one bundle hundreds of times.Accuracy — the material model scores mAP@50 0.69 on a held-out split of the same cameras (unseen moments). Classes with very few labels (solar 3, cable_reel 7, steel_rod_bundle 5) are not reliable ; more labelled frames are the only fix.
Material
DetectionCrowd ×6