Wildlife-Agentic-Vision

# Wildlife Behavioral Risk Vision Agent - Google Gemini 3 Hackathon

### Agentic Vision System using Google Gemini 3

An ethical, stateful, dual-agent wildlife monitoring system that detects **persistent behavioral deviations relative to group norms**, accumulates confidence over time, and visualizes at-risk animals using AI-seeded OpenCV tracking.

Built for conservation monitoring demos and the **Google Gemini 3 Hackathon**.

## Demo

<p align="center"> <img src="docs_assets/vision_demo.gif" width="400"> </p>

<p align="center"> <img src="docs_assets/sample.png" width="48%"> <img src="docs_assets/vision_sample.jpg" width="48%"> </p>

## Problem

Wildlife conservation teams often rely on manual observation to detect:

* Injured animals
* Social exclusion
* Persistent behavioral anomalies
* Abnormal movement patterns

This process is:

* Time consuming
* Subjective
* Not scalable

We built a **stateful agentic vision system** that:

* Observes animals over time
* Learns group behavioral baselines
* Detects persistent deviations
* Escalates risk without making medical claims


## System Overview

This project contains **two cooperating AI agents**:

### 1️⃣ Behavioral Reasoning Agent (Gemini 3 Pro)

Analyzes raw wildlife video and extracts structured behavioral signals:

* Stable `individual_id`
* Relative speed (0–1)
* Posture asymmetry (0–1)
* Distance from group center
* Group baseline averages
* Confidence accumulation
* At-risk seed (time window + normalized coordinates)

It outputs structured JSON memory for downstream reasoning.


### 2️⃣ Visualization Agent (Gemini 3 Pro)

Generates a complete standalone Python script that:

* Uses memory data as a **tracking seed**
* Seeks directly to the seed timestamp
* Converts normalized coordinates → pixels
* Initializes an OpenCV CSRT tracker
* Draws red bounding boxes for at-risk animals
* Saves annotated video output
* Captures sample frames


## Workflow

```
Wildlife Video
      │
      ▼
Gemini 3 Vision Agent
(Beh