AMD Slingshot — CCTV Surveillance & Abandoned Bag Detection

Real-time CCTV surveillance system for Indian railway stations that detects persons, bags, tracks ownership, and flags abandoned luggage using YOLO26 + BoT-SORT.

Architecture

┌─────────────┐    ┌──────────────┐    ┌─────────────────────┐    ┌──────────────────────┐
│  YOLO26m    │───>│  BoT-SORT    │───>│  OwnershipEvent     │───>│  AbandonmentRisk     │
│  Detection  │    │  Tracking    │    │  Engine             │    │  Engine              │
└─────────────┘    └──────────────┘    └─────────────────────┘    └──────────────────────┘
       │                  │                      │                          │
       │                  │                      │                          │
       │           ┌──────┴──────┐        ┌──────┴──────┐           ┌───────┴───────┐
       │           │ BagRegistry │        │  Ownership  │           │  AnomalyFusion│
       │           │ + ReID      │        │  FSM Logic  │           │  Engine       │
       └───────────┴─────────────┘        └─────────────┘           └───────────────┘

Core Components

Module	Description
run_ownership_tracking.py	Main pipeline — runs YOLO26m → BoT-SORT → ownership → abandonment
detection/detect.py	YOLO detection wrapper
detection/bag_registry.py	Appearance-based bag ReID with embedding similarity
detection/bag_reid_extractor.py	Feature extraction for bag re-identification
detection/ownership_event_engine.py	FSM-based person↔bag ownership tracking
detection/ownership_manager.py	Legacy IoU + co-movement ownership logic
detection/abandonment_risk_engine.py	Multi-factor abandonment risk scoring
detection/anomaly_fusion_engine.py	Fuses anomaly signals from multiple engines
detection/association.py	Hungarian algorithm person-bag association
infer_anomaly.py	Standalone anomaly inference script
botsort.yaml	BoT-SORT tracker configuration

Dataset Pipeline

Tools for building a fine-tuning dataset from raw CCTV footage:

Script	Purpose
extract_frames_hd.py	Extract full-resolution frames (1280×720) from MP4 videos
sample_frames.py	Fixed-interval sampling with optional perceptual dedup
sample_motion.py	Motion-aware sampling using histogram difference (best for static CCTV)
validate_dataset.py	Pre-training dataset validator (format, balance, duplicates, tiny boxes)

Dataset Workflow

# 1. Extract HD frames from video
python extract_frames_hd.py --src . --out frames_hd --interval 2

# 2. Motion-aware sampling (picks diverse frames)
python sample_motion.py --src frames_hd --out dataset --target 3000

# 3. Auto-label with COCO model
yolo predict model=yolo26m.pt source=dataset/images/train imgsz=1280 conf=0.15 save_txt=True classes="[0,24,26,28]"

# 4. Review labels in LabelImg
labelImg dataset/images/train dataset/classes.txt dataset/labels/train

# 5. Validate before training
python validate_dataset.py --data dataset

# 6. Fine-tune
yolo train model=yolo26m.pt data=dataset/dataset.yaml epochs=30 imgsz=1280 batch=4 amp=True

Detection Classes

ID	Class	Use Case
0	person	Track owners, detect crowds
1	backpack	Most common bag type in railways
2	handbag	Smaller personal bags
3	suitcase	Large luggage on shelves

Tracking Configuration

• Model: YOLO26m (detection, not classification)
• Tracker: BoT-SORT (botsort.yaml)
• Resolution: 1280px inference
• Confidence: 0.25 (tracking) / 0.15 (auto-labeling)
• ReID: Handled by BagRegistry (BoT-SORT ReID disabled)

Hardware

• GPU: NVIDIA RTX 4060 (8GB VRAM)
• Training: batch=4, imgsz=1280, amp=True
• Inference: ~50ms/frame at 1280px

Environment

Camera	Location	Characteristics
ADMIN Block	Chemical Corridor, C-119	Indoor, fluorescent lighting
Admin Block 2	Near B301	Indoor, corridor
CSED	Department entrance	Low contrast, haze
Library 2	Baggage Area	Shelf-packed luggage, dark-on-dark
Main Gate	Entry/Exit	Outdoor, high traffic
Parking	Learning Block	Outdoor, varying light
Road	Hostel-H Chownk	Outdoor, 3200×1800 resolution

Requirements

pip install -r requirements.txt

Key dependencies: ultralytics, opencv-python, torch, numpy, scipy, imagehash, pillow, tqdm

License

Private repository — all rights reserved.