Project V1.2 — Multi-Camera People Tracking with Cross-Camera Re-Association and Bird's-Eye-View Spatial Mapping

Overview

An end-to-end multi-camera surveillance pipeline that maintains consistent person identities across 3 synchronized camera views and projects real-time positions onto a bird's-eye-view floor map. Built on the EPFL Multi-Camera Pedestrian dataset (Laboratory sequence, 4 people, 500 frames per camera).

Target roles: Video Analytics | Smart City Surveillance | Multi-Camera Systems

---

Demo

Composite Output — 3 Camera Feeds + BEV Map

Three synchronized camera feeds with global ID overlays (L: local ID, G: global ID) alongside real-time bird's-eye-view trajectory animation.

(embed composite.mp4 or GIF here)

BEV Map Animation

Color-coded person trajectories on a top-down floor plan with 30-frame trajectory tails.

(embed bev_animation.mp4 or GIF here)

---

System Architecture

Stage	Component	Detail
Detection	YOLOv8m	MPS inference, person class only
Tracking	StrongSort (boxmot v16)	Per-camera local track IDs
Embedding	OSNet x1_0	512-d L2-normalized appearance embeddings
Association	Cosine Similarity + EMA	Cross-camera global ID assignment
Projection	cv2.perspectiveTransform	EPFL calibration homographies
Visualization	OpenCV	BEV animation + composite MP4 export

---

Results

Metric	Tuned (visual)	Original
CMC@1	50.0%	62.5%
CMC@3	90.0%	87.5%
CMC@5	90.0%	87.5%
mAP	56.8%	62.6%

Note: Threshold tuning (0.75 → 0.60) with second-pass prototype merging improves cross-camera visual consistency at the cost of CMC@1. CMC@3 remains strong at 90.0%, indicating correct matches are consistently found within top-3 ranks. The primary failure mode with pretrained weights is extreme viewpoint change (frontal vs. rear appearance), addressable via domain-specific fine-tuning.

---

Dataset

EPFL Multi-Camera Pedestrian — Laboratory Sequence

• 4 synchronized cameras, 4 pedestrians, 500 frames @ 25fps
• Provided homography calibration used directly for BEV projection
• Source: https://www.epfl.ch/labs/cvlab/data/data-pom-index-php/

---

Setup

conda create -n v12_multicam python=3.10 -y
conda activate v12_multicam
pip install torch torchvision torchaudio
pip install ultralytics boxmot opencv-python numpy scipy pyyaml tqdm matplotlib torchreid tensorboard

Download EPFL Laboratory sequence videos into data/raw/:

curl -L -o data/raw/cam0.avi "https://documents.epfl.ch/groups/c/cv/cvlab-pom-video1/www/4p-c0.avi"
curl -L -o data/raw/cam1.avi "https://documents.epfl.ch/groups/c/cv/cvlab-pom-video1/www/4p-c1.avi"
curl -L -o data/raw/cam2.avi "https://documents.epfl.ch/groups/c/cv/cvlab-pom-video1/www/4p-c2.avi"

---

Run

# 1. Extract frames
python src/extract_frames.py

# 2. Parse EPFL calibration homographies
python -m src.parse_calibration

# 3. Run per-camera detection + tracking + embedding
python -m src.pipeline_single_cam --cam cam01
python -m src.pipeline_single_cam --cam cam02
python -m src.pipeline_single_cam --cam cam03

# 4. Cross-camera ReID association
python -m src.run_association

# 5. Full pipeline — BEV animation + composite export
python -m src.main_pipeline

# 6. ReID evaluation
python -m src.eval_reid

---

Environment

• M1 MacBook Air — all pipeline work (detection, tracking, embedding, BEV)
• Device: MPS for YOLOv8m and OSNet, CPU fallback for StrongSort
• Resolution: 640px | num_workers=0 throughout

---

Key Learnings

• Cross-camera ReID with cosine similarity and EMA prototype updates
• Homography-based ground plane projection from EPFL calibration files
• Two-pass global ID merging to resolve viewpoint-induced identity splits
• Precision-recall tradeoff between CMC@1 and visual tracking consistency

---

Project Roadmap

This is Project V1.2 in a video analytics portfolio series:

• V1.1 — Multi-Object Tracking with StrongSORT + ByteTrack (MOT17, HOTA evaluation)
• V1.2 — Multi-Camera Tracking with Cross-Camera ReID + BEV Mapping ← you are here