Mohit Patel

06 · Edge AI · Computer vision · Geospatial · Research

Real-Time Wildlife Monitoring

Research paper: “Adaptive Occlusion Edge Intelligence for Real-Time Wildlife Monitoring and Water Source Analysis Across Diverse Ecosystems” - Mohit Patel, Jabalpur Engineering College

An edge-AI system on sub-$150 hardware that detects, tracks and maps wildlife around water sources - 87.3% mAP@0.5 at 7.7 FPS on a Raspberry Pi 4, with an occlusion-aware method that adds 11.2 points under forest canopy.

87.3%
mAP@0.5 (15 species)
7.7 FPS
Raspberry Pi 4 · 5.9 W
+11.2 pts
Under heavy canopy

How we built it

  1. The problem: water sources decide where wildlife gathers, especially in summer, but collars, camera traps and manual surveys are slow, costly and intrusive.
  2. Ground nodes: a Raspberry Pi 4 with Camera Module v2 and a Neo-6M GPS (5-8 m accuracy) runs the full pipeline at the water hole; ESP32-CAM units wake on a PIR sensor and take one image every 5 s at 0.8 W.
  3. Aerial survey: drone imagery over a 2 km² mixed-deciduous forest plot in central India - 50 m altitude was chosen after testing 30, 50 and 70 m. My own 2.5-inch FPV drone carries a Raspberry Pi and a detection camera for low passes over water holes.
  4. Detection: YOLOv8n fine-tuned on 12,000 hand-annotated frames of 15 species (forest plot, drone and Snapshot Serengeti), quantised to INT8 - 89 ms per frame on the Pi.
  5. My occlusion-aware confidence modulation scores how hidden an animal is from image texture and edges, and boosts confidence near water: under heavy canopy, mAP rose from 75.8% to 87.0%.
  6. Tracking: Deep SORT with an occlusion penalty keeps each animal’s ID - 55% fewer ID switches (4.1 vs 9.2 per 1,000 frames) and MOTA up from 73.8% to 79.1%.
  7. Mapping: every detection is GPS-tagged and pinned on the map; kernel-density heatmaps with 50/75/90/95th-percentile contours reveal saturation zones - 124 m mean radius around the main water source, from 142 m for chital to 85 m for leopard.
  8. Counting: animals are tallied per species and saved for the forest administration, so they know how many of each species use each water source.
  9. Results: 87.3% mAP@0.5 overall, 5.7 points above MegaDetector v5; 90.4% species classification; whole pipeline 130 ms per frame (7.7 FPS) at 5.9 W.

Hardware & tools