A Real-Time Edge-Based Deep Learning Framework for Intelligent Home Intrusion and Object Removal Detection

1. Problem Statement

The transition toward intelligent home monitoring is hindered by the limitations of traditional motion-triggered systems, which lack the discriminative capacity to distinguish between benign movement (e.g., pets, swaying curtains) and malicious activities [1]. While cloud-based deep learning solutions offer high accuracy, they introduce unacceptable latency—often exceeding the critical window for intrusion response—and raise substantial privacy concerns regarding the transmission of private household video streams to third-party servers [2], [8]. Moreover, a significant gap exists in the detection of "Object Removal." Standard intrusion systems focus on boundary crossing but fail to detect when a specific high-value asset is displaced or stolen by an individual who may have initially bypassed detection. There is a lack of unified frameworks that can simultaneously verify identity, track object persistence, and analyze behavioral patterns locally on resource-constrained edge devices [14]. This research addresses these challenges by proposing a resource-efficient, multi-tasking architecture capable of performing real-time semantic analysis at the network edge.

2. Literature Review and Research Gaps

The field of intelligent surveillance has evolved through three distinct phases: traditional background modeling, cloud-based deep learning, and current edge-centric architectures.

2.1. Motion Detection and Object Tracking

Early systems relied on Gaussian Mixture Models (GMM) and frame differencing. While computationally inexpensive, these methods are highly sensitive to illumination fluctuations and lack semantic depth [9]. Subsequent advancements in Convolutional Neural Networks (CNNs) significantly improved detection accuracy but were initially too heavy for real-time deployment on embedded systems [11].

2.2. Edge Intelligence and Model Compression

To enable edge deployment, research has shifted toward model compression. Techniques such as weight pruning and quantization-aware training have been shown to reduce the footprint of models like YOLO and Mobile Net without drastic accuracy degradation [4]. However, most studies focus on generic object detection rather than the specific temporal logic required for intrusion and theft detection [18].

2.3. Action Recognition and Anomaly Detection

Action recognition using CNN-LSTM architectures has shown promise in identifying suspicious behavior [13]. However, these models often suffer from high computational latency, making them impractical for low-power edge gateways. Furthermore, the "black-box" nature of these models poses a challenge for residential users who require justification for triggered alarms [5].

2.4. Identified Research Gaps

1. Semantic Object Persistence: Existing edge systems rarely correlate person identity with the state of specific objects in the environment.
2. Contextual Anomaly Scoring: Most systems use binary classification, ignoring the nuanced spatial-temporal trajectories that distinguish an intruder from a resident [12].
3. Privacy-Preserving Explainability: There is a dearth of research into providing visual and feature-based explanations for anomalies on edge devices without compromising raw data privacy [6].

3. Proposed Methodology

The proposed framework follows a hierarchical execution strategy designed to minimize power consumption by activating high-complexity modules only when necessary.

3.1. Edge Processing Pipeline

The system operates via a continuous "Sensing and Trigger" loop:

1. Lightweight Ingestion: Frames are captured and normalized.
2. Primary Trigger (Person Detection): A pruned YOLO-Edge model scans for human presence.
3. Secondary Verification: If a person is detected, the system branches into (a) Face Authentication to verify residency and (b) Object-State Comparison to check the status of high-value assets.
4. Tertiary Analysis (Behavioral Scoring): If the identity is unverified or "Unknown," the system tracks the subject’s trajectory and computes an anomaly score Ψ

   

5. Explainability and Alerting: For high-score anomalies, the system generates Grad-CAM heatmaps and SHAP importance values, transmitting only the metadata and the XAI-enhanced frame to the user [7].

3.2. Model Optimization for Edge Devices

We implement a three-tier optimization strategy:

• Backbone Selection: We utilize a MobileNetV3-Small backbone for the detection head to reduce the parameters by 60% compared to standard YOLO architectures [11].
• Global Structured Pruning: Layers with low L2

  

  -norm activation are pruned, followed by fine-tuning to recover accuracy.
• Tensor RT Quantization: The model is converted to INT8 precision using a calibration dataset that mimics varied lighting conditions [4].

4. System Architecture

4.1. Person Detection Module (PDM)

The PDM is the "always-on" component. It utilizes a modified YOLOv8-Nano architecture optimized for indoor aspect ratios. It identifies the bounding box Bp

4.2. Face Authentication Module (FAM)

Once Bp

This vector is compared against a local database of authorized embeddings Vauth

4.3. Object-State Comparison Module (OSCM)

The OSCM focuses on "Regions of Interest" (RoI) where high-value objects (e.g., safes, laptops) are located. It employs a persistence counter for each object Oj.

4.4. Behavior-Aware Anomaly Detection Module (BAAD)

The BAAD module analyzes the sequence of centroids C={c1,c2,...,ct}

over a temporal window W. It uses a lightweight GRU to predict the next likely position; significant deviations from "Routine Pathing" increase the anomaly score [17].

4.5. Alert and Logging Subsystem

To ensure privacy, the subsystem converts detections into encrypted JSON metadata. Raw video is stored in a rolling 24-hour buffer locally, with only XAI-stamped frames uploaded to the user upon a confirmed alert.

5. Mathematical Modeling

5.1. Intrusion Detection Logic

Let Z

 at time t is modeled as: 

where 1

5.2. Object Persistence Modeling

The state of a monitored object Oj

We model the "Removal" event through a temporal decay function:

where Mref

5.3. Behavioral Anomaly Scoring

The behavioral score Ψ is calculated based on the entropy of the path H(C) and the velocity V  of the subject:

where 

5.4. Performance Evaluation Metrics

We evaluate the framework using the F1-score and the Latency-Accuracy Trade-off (LAT):
 

LAT=Precision⋅RecallInference Latency (ms)

This metric emphasizes the necessity for both high detection rates and rapid edge execution [18].

6. Experimental Setup

6.1. Edge Configuration

All experiments were conducted on an NVIDIA Jetson Orin Nano (8GB RAM) utilizing Tensor RT 8.6 and CUDA 11.4. The camera input was a 1080p stream at 30 FPS.

6.2. Custom Dataset Assumptions

We utilized a custom residential dataset consisting of 12,000 frames. The dataset incorporates:

• Class Imbalance: 85% normal household activity, 10% simulated intrusions, 5% object removal events.
• Environmental Variability: Scenes include day, night (IR), and sudden illumination changes.
• Simulated Anomalies: Intruders wearing masks, loitering behavior, and rapid asset removal [16].

7. Comparative Analysis

The framework was benchmarked against four baseline architectures: