1 of 9

HONOURS PROJECT

Multi-Modal Behavioral

Indexing for Edge-Based

Threat Detection

Jayden-Lee Van Rooyen · 4069341 · University of the Western Cape

Edge Computing

Sensor Fusion

Privacy-by-Design

Behavioral Analytics

Time-of-Flight

2 of 9

THE PROBLEM

02

Reactive, Not Proactive

Traditional CCTV systems serve as forensic tools for post-event analysis. Manual monitoring cannot scale to modern video volumes.

Cloud Dependency

Existing AI approaches stream raw video to centralised servers, introducing latency, bandwidth costs, and significant privacy risk.

📐

No Depth Perception

Vision-only systems infer proximity from bounding-box size which is an estimate that degrades under occlusion and perspective distortion.

This project addresses all three by fusing RGB vision + ToF depth on an autonomous edge node with no cloud required.

3 of 9

BACKGROUND & RELATED WORK

03

01

Visual Perception

Gap: 2D only (no metric depth)

Sapkota, Zheng, Wojke, Mandalapu

02

Sensor Fusion

Gap: Assumes server-class hardware

Gimpelj, Park, Romanenko, Li & Tang

03

Motion & Pose

Gap: No simultaneous depth pairing

Wahyono, Mai & Bui, Oh et al.

04

Privacy & Behavior

Gap: No unified numeric output

Song, Al-Madani, Qamar

05

Threat Indexing

Gap: Pixel-space proximity proxy

Clarke, Pouyan, Wu et al.

06

Edge Deployment

Gap: Single-stream only (no fusion)

Pereira, Reis, Ali, Chen

No existing system unifies all six pillars into a single privacy-preserving edge deployment, this is the research gap.

4 of 9

THE BEHAVIORAL THREAT INDEX

04

T = (w₁ · Temporal) + (w₂ · Spatial) + (w₃ · Behavioral)

⏱ w₁ - TEMPORAL

Dwell-Time Tracking

  • Persistent Tracking IDs via DeepSORT
  • Loitering defined as spatio-temporal accumulation
  • 45% reduction in identity switches during occlusion

Wahyono et al. [18], Wojke et al. [19]

📡 w₂ - SPATIAL

Depth Zone Detection

  • 8x8 VL53L5CX ToF distance matrix at I2C
  • Sterile zone breach detection with metric accuracy
  • 63° FoV aligned to Camera Module 3 via calibration

Gimpelj & Munih [5], Romanenko [15]

🦾 w₃ - BEHAVIORAL

Anomaly Scoring

  • Skeletal keypoint extraction via Hailo-10H NPU
  • mSSA trajectory outlier detection
  • 96.51% action classification accuracy (Mai & Bui)

Mai & Bui [7], Oh et al. [9]

Weights determined via AHP-Entropy-CRITIC framework (Wu et al. [20])

5 of 9

SYSTEM HARDWARE

05

Raspberry Pi 5

8GB · 2.4GHz Quad-core A76

Role: Host Logic / T Calculation

Hailo-10H NPU

AI HAT+ · 40 TOPS · PCIe 3.0

Role: Inference for w₁, w₃

Camera Module 3

12MP / HDR / PDAF · CSI-2

Role: Visual Input (x₁, x₃)

VL53L5CX ToF

8×8 Multi-zone · I2C Bus

Role: Proximity Input (x₂)

27W USB-C PSU

Raspberry Pi Official

Role: System Stability

Active Cooler

Variable Speed Fan

Role: Prevent Thermal Throttling

All inference runs locally on-device. No raw video is transmitted or stored at any point.

6 of 9

THREAT MODEL & SECURITY SCOPE

06

👁

Physical Tampering

Intruder covers lens or ToF matrix to induce sensor blindness and prevent detection.

🎭

Adversarial Attacks

Physical adversarial patches suppress human-class detection within the YOLO pipeline (w₃).

📶

Data Integrity

I2C bus transmission exposes ToF readings to potential spoofing, causing false negatives in w₂.

🔓

Privacy Leakage

Unauthorized physical access could allow reconstruction of retained behavioral metadata.

Privacy-by-Design: Raw frames purged from RAM immediately after keypoint extraction. Only Tracking IDs, depth coords, and T transmitted.

7 of 9

SYSTEM REQUIREMENTS

07

FUNCTIONAL REQUIREMENTS

FR-1

Multi-Modal Data Ingestion

1080p RGB via CSI + 8×8 depth via I2C concurrently

FR-2

Edge-Based Pose Estimation

NPU accelerator extracts skeletal keypoints in real-time

FR-3

Late-Fusion Spatial Mapping

2D bounding boxes mapped to 3D ToF coordinates

FR-4

Threat Index Calculation

Composite T computed every 200ms via weighted sum

FR-5

Automated Alert Triggering

Local notification when T exceeds safety threshold

NON-FUNCTIONAL REQUIREMENTS

Performance

< 150ms latency · ≥ 20 FPS · < 75% CPU

🔒

Security & Privacy

No raw video stored or transmitted

📈

Scalability

Additional ToF sensors via I2C multiplexer

🖥

Usability

Admin adjusts wₙ weights without coding

🔧

Maintainability

Virtual env isolation + comprehensive logging

8 of 9

SYSTEM PIPELINE & IMPLEMENTATION CHALLENGES

08

1

CAPTURE

RGB 1080p + 8×8 ToF

2

INFER

Hailo-10H NPU

Pose Estimation

3

FUSE

Late Fusion

Spatial Mapping

4

SCORE

Compute T

every 200ms

5

ALERT

Local notification

if T > threshold

🔒 Raw frames purged from RAM immediately after Step 2 so that only anonymized metadata persists.

Calibration

Aligning the 63° FoV of the VL53L5CX with Camera Module 3 so depth data maps accurately to visual bounding boxes. The asynchronous data rates are the central engineering challenge.

Simulation

Using mSSA (multidimensional Singular Spectrum Analysis) to test trajectory outlier detection and index weight sensitivity in a controlled virtual environment before hardware deployment.

Integration

Implementing the Hailo-10H Python pipeline to evaluate real-time inference latency and thermal stability under sustained 24/7 load on the Raspberry Pi 5.

9 of 9

ROADMAP (TERM 2)

09

WHAT WE PROPOSE

PHASES

🔧

PHASE 1: Hardware assembly & sensor mounting

📐

PHASE 1: ToF-to-Camera FoV calibration

💻

PHASE 2: Hailo-10H NPU pipeline implementation

🧪

PHASE 2: mSSA simulation & index weight tuning

📊

PHASE 3: Real-time latency & thermal benchmarking

Goal: Prove that sophisticated behavioral analytics can run entirely at the edge while preserving privacy without sacrificing accuracy.