EffiPed

Multi-Camera Pedestrian Detection, Tracking & Re-Identification using Joint ConvNeXt V2 Architecture

By Aswanth Raj

EffiPed is a compact video-intelligence system that detects pedestrians, maintains camera-local tracks, and ranks cross-camera identity candidates for human review. Its React investigation console is available as a precomputed browser demo; the same workflow can connect to local FastAPI/CUDA inference when an authorized checkpoint is available.

Important

Ranked matches are reviewable appearance evidence, not proof of identity. The hosted experience is a precomputed, non-commercial research demonstration. Public model weights remain withheld while training-data redistribution terms are unresolved.

Try the demo workbench

The hosted UI mirrors the original PedestrianTracker application layout, with four P-DESTRE session 12-11-2019_3 clips already attached as though you had uploaded them. Six panels:

• Single Camera — per-camera tracked output and run summary;
• Cross Camera — four-view association replay plus the per-pair precision table;
• Person Search — the main feature: build the index, pick any detected person, and review their stored appearances and ranked cross-video candidates;
• Image Detection — a single-frame detection pass;
• Model Status and Research Context — runtime and protocol notes.

cd apps/web
npm install
npm run dev

The controls are live, but the hosted build performs no inference: each Run replays a precomputed result. Full-frame views are redrawn in the browser by seeking the shipped clip to the appearance's timestamp and stroking its stored box, so no per-appearance scene images ship.

No synthetic browser boxes are drawn over the footage. The boxes baked into the replay videos are the annotations rendered by the original tracking pipeline.

Note

The person-search index was computed offline with the BoxJDE research checkpoint, because the EffiPed Tier-1 weights are withheld pending dataset rights review. The Person Search and Model Status panels both state this. Regenerate it with tools/build_person_search_fixture.py.

System

One ConvNeXt V2 feature hierarchy supports CenterNet-style detection and a 256-D part-aware identity descriptor. RoIAlign extracts a person feature map, four horizontal body strips retain local appearance, and Coordinate Attention fuses the visible evidence. BoT-SORT combines motion, overlap, and appearance for local temporal association; the gallery then ranks possible cross-camera matches for an analyst.

Evaluation	Result
P-DESTRE validation cross-camera Rank-1	62.8%
P-DESTRE test cross-camera Rank-1	61.3%
P-DESTRE validation / test detection mAP@0.5	90.74% / 88.4%
MOT17 val-half MOTA / IDF1 / HOTA	64.08 / 74.24 / 61.34
EffiPed Tier-1 footprint	7.78M · ≈18 full-pipeline FPS

Each value has a protocol label in RESULTS.md. The interactive replay is an application demonstration, not a benchmark run.

Architecture

The diagram is also available as an editable PowerPoint.

Repository map

src/effiped/          installable model, descriptors, tracking, runtime
apps/api/             FastAPI local-GPU service and job lifecycle
apps/web/             React/Vite demo workbench UI and hosted replay
configs/system/       active EffiPed and matched PartJDE configurations
research/results/     single source of truth for published evidence
research/report/      generated technical report
docs/architecture/    editable diagram source and web exports
docs/media/           optimized, attributed demonstration media
tools/ and tests/     validation, regression, and release checks

Run live inference locally

Python 3.11 and an NVIDIA GPU are recommended.

python -m venv .venv
# Windows: .venv\Scripts\activate
# Linux/macOS: source .venv/bin/activate
pip install -e ".[runtime]"
effiped-app

Place an authorized checkpoint in EFFIPED_WEIGHTS_DIR. When none is present, the API reports the model as unavailable without exposing a local filesystem path.

effiped-train --config configs/system/effiped-tier1.yaml
effiped-eval --config configs/system/effiped-tier1.yaml
effiped-demo

Variable	Purpose
EFFIPED_WEIGHTS_DIR	authorized local model artifacts
EFFIPED_RUNTIME_DIR	temporary uploads, crops, and job assets
EFFIPED_DEVICE	auto, cpu, cuda, or cuda:N
EFFIPED_MAX_UPLOAD_MB	per-video upload limit
EFFIPED_ALLOWED_ORIGINS	comma-separated CORS allowlist

Public API

• GET /api/health
• GET /api/models
• POST /api/person-search/jobs
• GET /api/person-search/jobs/{job_id} and /stream
• GET .../people, /detections, /tracks, and /matches
• POST .../search-by-example
• DELETE /api/person-search/jobs/{job_id}
• GET /api/assets/{asset_id}

Deleting a job removes uploaded video and generated assets.

Research connections

The later BoxJDE Person Search repository isolates the full-person descriptor readout and documents its five-fold P-DESTRE ablation. It is linked as related research; its code and report are not duplicated here.

Licensing and responsible use

Original software is © 2026 Aswanth Raj and licensed under Apache-2.0. P-DESTRE-derived media under docs/media/pdestre/ is separately licensed as a CC BY-NC-SA 4.0 adaptation for this non-commercial showcase. The asset manifest records the source, transformations, hash, purpose, and license for every derived asset.

No dataset, source video, person-level benchmark record, checkpoint, or runtime crop is included.

Model card · Data and weight-release audit · Third-party notices · P-DESTRE paper · CC BY-NC-SA 4.0