Research Area 4 · Visual Memory & Scene Comparison

Research 4: Visual Memory & Scene Comparison

Investigating interval-based change detection, structural gradient matching for illumination/shadow invariance, and lightweight edge-AI vs. cloud-dependent architectures.

Research Question 4

How can an AI camera system spot real changes in a room (like a moved or missing object) every 10 seconds without being fooled by shadows or light changes?

Specific Problem Statement & Strategy

Conventional security cameras often produce false alarms from shifting sunlight, moving shadows, and transient ambient motion, while struggling to reliably detect permanent, meaningful changes in physical room state (such as moved, missing, or newly introduced equipment).

🖥️ Edge Compute

Local Mac mini / Linux host processing eliminates recurring cloud compute and API subscription costs.

📹 1080p Dual Cameras

Two synchronized 1080p cameras provide cross-viewpoint redundancy and eliminate single-camera blind spots.

⏱️ 10–15s Check Interval

Captures static frame pairs periodically instead of streaming 4K continuously, minimizing CPU and thermal load.

🗑️ Prune Static Frames

If algorithmic comparison reveals no significant visual delta against the baseline, prior static frames are purged locally.

Scholarly Research & Academic Baseline

Analysis of recent peer-reviewed change-detection architectures and visual memory frameworks:

Source / Paper	Core Mechanism	Key Difference & Jarvis Edge Gap
Lu, Zhou, & Chen (2026) GCN-Mamba Remote Sensing Change Detection	Combines Graph Convolutional Networks (GCN) and Mamba state-space models with local-global feature aggregation to evaluate frame pairs and filter lighting shifts.	Designed for high-altitude top-down satellite imagery with massive cloud compute. Jarvis adapts local-global feature aggregation for ground-level 1080p indoor perspectives on local edge hardware.
Traffic Vision Group (2026) UAV Traffic Scene Understanding with TRM	Employs a Traffic Regulation Memory (TRM) module to track spatial states over time and discard redundant visual noise without continuous raw video streaming.	Targets aerial drone traffic with heavy multi-modal cloud processing. Jarvis applies interval visual memory to a lightweight edge setup (Mac mini / Linux) optimized for indoor spaces.

Consolidated Synthesis: Academic research confirms that feature-aggregation models coupled with persistent visual memory reliably isolate genuine structural changes from lighting noise. Jarvis fills the gap by packaging these principles into a standalone, edge-based framework operating at discrete 10-second intervals.

Patent Research & Prior IP

Patented machine vision techniques addressing illumination invariance and background subtraction:

Patent Reference	Patented Innovation	Relevance & Implementation Tradeoff
US7062093B2 Steger (2006) · Illumination & Shadow Invariance	Utilizes structural gradient shapes and edge-based matching rather than raw pixel intensities, achieving invariance against dynamic shadows, contrast reversals, and lighting switches.	Provides the exact mathematical basis to prevent shadow false alarms. However, full-frame continuous gradient matching is computationally heavy; Jarvis bounds edge-matching to detected object regions.
US6678413B1 Liang et al. (2004) · Object Tracking & State Monitoring	Couples video capture with automated background subtraction to establish a baseline model and register displaced or missing assets over time.	Directly aligns with Jarvis's missing-asset tracking goal. Unlike Liang's continuous high-frame-rate approach, Jarvis executes baseline subtraction at 10-second intervals to minimize storage.

Commercial Products & Competitive Analysis

Evaluation of existing consumer and enterprise camera systems:

Commercial System	Core Features	Key Limitations	How Jarvis Improves
Ring Stick Up Camera Smart Home Security	HD streaming, motion triggers, and periodic "Snapshot Capture" intervals uploaded to the cloud.	Cloud lock-in, recurring subscription fees, storage bloat, and naive pixel-motion triggers that lack semantic object understanding.	100% Local & Private: Jarvis processes frames on-device, drops redundant static captures, and evaluates semantic changes without subscription fees.
Viewtron AI Camera & NVR Commercial NVR Security	Configurable "Missing Object" polygon boundary zones with integrated NVR alert hardware.	Bulky proprietary NVR hardware, heavy multi-cable wiring, continuous stream processing, and high deployment cost.	Lightweight Edge Setup: Jarvis replaces bulky NVR appliances with a standalone edge machine executing interval checks every 10 seconds.

Ring Stick Up Camera

Cloud-dependent snapshot model

Viewtron AI NVR System

Bulky commercial NVR appliance

Project Team Arnav Mittal, Shriya Paladagu, Soni Dhenuva, Xavier Thompson

Prototype Stage View Repo