🐘 EleGuard: Multimodal Elephant Detection

EleGuard is an intelligent wildlife surveillance system designed for the 24/7 monitoring and detection of elephant activity in natural habitats. This project is a submission for the Gemma 4 Good Hackathon and focuses on local, offline deployment to mitigate human-elephant conflict in rural communities across Sri Lanka.

EleGuard is a model trained on a dataset based on Gemma 4 E2B.

Features

• Multimodal Detection: Integrated visual (Infrared/Daytime) and acoustic analysis for robust monitoring.
• Reasoning Distillation: Utilizes expert logic derived from Gemini 3.1 Flash to provide contextual safety alerts.
• Edge-Optimized: Built for offline, CPU-only execution on standard hardware using llama.cpp.
• Low-Connectivity Ready: Designed for resource-constrained environments with no cloud dependency.

Technical Architecture

The system uses a Teacher-Student Knowledge Distillation framework to achieve high intelligence in a small footprint.

1. The Teacher Model: Gemini 3.1 Flash analyzed 2,600 samples to generate detailed "thought blocks" explaining visual and acoustic features.
2. The Student Model (EleGuard): Based on Gemma 4 E2B, fine-tuned using the Unsloth framework to internalize this expert reasoning.

EleGuard Dashboard

The dashboard provides a real-time interface built with Streamlit.

• Multi-Camera Support: Simultaneous monitoring of multiple video feeds.
• Motion Filtering: OpenCV-based tracking and camera shake suppression.
• Local Inference: Connects to a local llama.cpp server for offline processing.

How It Works

To ensure reliable, 24/7 monitoring without overwhelming edge hardware, EleGuard is structured as an event-driven inference cascade. By using lightweight heuristics to filter raw environmental data, the system ensures the compute-intensive Gemma 4 reasoning model is only triggered during high-probability events. The complete pipeline, from sensor capture to local alert generation, is illustrated below.

Installation and Usage

1. Clone the repository:

git clone [https://github.com/MalithaBandara/EleGuard.git](https://github.com/MalithaBandara/EleGuard.git)
cd EleGuard

2. Install dependencies:

pip install opencv-python streamlit librosa numpy moviepy requests

3. Download llama.cpp

To run the local inference server, you need the llama.cpp binaries. It is recommended to use build b9090 or newer for full Gemma 4 multimodal support.

1. Go to the llama.cpp Releases page.
2. Look for Release b9090 (or the latest version).
3. Download the specific zip for Windows CPU: llama-b9090-bin-win-cpu-x64.zip.
4. Extract the contents of the zip file.

4. Download Model Weights (GGUF)

Download the following required files from the EleGuard Hugging Face Repository:

• EleGuard-gemma-4-E2B-it.Q4_K_M.gguf (Main Model)
• EleGuard-gemma-4-E2B-it.F16-mmproj.gguf (Vision Projector)

Place these files in your project root or a known directory (e.g., C:\EleGuard).

5. Start the Local Inference Server

Run the llama.cpp server on port 8080.
Note: You must update the three paths in the command below to match the actual locations on your system (the server executable, the model file, and the vision projector).

# Ensure you update the paths below to point to your local files
C:\EleGuard\llama-b9090-bin-win-cpu-x64\llama-server.exe -m C:\EleGuard\gemma-4-e2b-it.Q4_K_M.gguf --mmproj C:\EleGuard\gemma-4-e2b-it.F16-mmproj.gguf -c 512 -ngl 0 --host 127.0.0.1 --port 8080

6. Launch the Dashboard:

python -m streamlit run app.py

Configuration

EleGuard allows important runtime parameters to be customized through the config.json file.

Parameter	Description
inference_resolution	Resolution used for local model inference. Lower resolutions improve speed and reduce CPU usage.
jpeg_quality	JPEG compression quality applied before sending images for inference. Lower values reduce image size and latency.
inference_interval	Time interval (in seconds) between inference runs.
min_object_area_fraction	Minimum detected motion area required to trigger analysis. Helps filter small movements and noise.
persistence_required	Number of consecutive detections required before triggering inference. Reduces false positives.
max_fg_ratio	Maximum foreground motion ratio allowed before treating the frame as camera shake or excessive movement.
max_scatter_contours	Maximum number of scattered motion contours allowed before ignoring the frame.
audio_threshold	Minimum audio activity threshold required for audio-triggered analysis.
alert_cooldown	Cooldown period (in seconds) between alerts to prevent repeated triggering.
model_url	URL endpoint used for local Gemma inference through llama.cpp.
max_tokens	Maximum number of generated response tokens from the model.
temperature	Sampling temperature controlling response randomness and determinism.

This configuration system allows EleGuard to be adapted for different:

• camera setups
• environmental conditions
• hardware capabilities
• deployment scenarios
• latency requirements

Impact and future work

EleGuard serves as a practical blueprint for localized conservation technology. It demonstrates that multimodal AI models such as Gemma 4 can be compressed, optimized, and deployed locally to solve real-world safety problems.

The system is fully open source, allowing others to inspect, improve, and adapt it for real-world deployment. Future versions could integrate low-power edge devices, SMS gateways, LoRaWAN-based mesh networks, and solar-powered deployments for fully autonomous monitoring.

Resources

• Model Weights (GGUF): Fine-tuned weights and multimodal projectors are hosted on Hugging Face.
• Dataset: The model was trained on the EleGuard Dataset, which contains 2,600 curated samples of infrared/daytime imagery and bioacoustic spectrograms.

Legal and Trademarks

• Gemma is a trademark of Google LLC.
• EleGuard is a model trained on a dataset based on Gemma 4 E2B.
• This project is independently developed and is not an official Google release.