← Back to DIY

Field build log

My Edge AI Wildlife Observer

I built a tiny field scientist that watches the world without uploading everything to the cloud. Jetson Nano for on-device classification, solar for power, and an SD card for data. No SIM card, no subscription, no latency.

Build time

3 weekends

Core brain

Jetson Nano

Power

Solar + SLA

Forest trail with dappled light — the kind of spot you would place a camera trap

01 / Plan

Design and constraints

I wanted a tiny field scientist that watches the world without uploading everything to the cloud. The Jetson Nano classifies locally, stores only confirmed sightings, and runs on solar.

The core constraint was power. A Jetson Nano draws 5-10W — which is a lot for a solar-powered device that runs 24/7. I designed the entire system around sleep cycles.
Classification happens on-device using a MobileNet-SSD variant fine-tuned on local wildlife. No cloud dependency means no cellular bill and no privacy concerns.
The housing must survive rain, humidity, insects, and curious animals. I used an IP67 junction box with a silica gel pack inside.

Click to explore

Inside the observer

IP67 ENCLOSUREGPUCPU20W SOLAR12V 7AhPIRSD

Selected component

Jetson Nano

Edge AI brain

Runs inference on the 128-core Maxwell GPU. Classifies species locally in ~200ms. No cloud required.

Field note: Thermals matter in an enclosed box. Add thermal pads to the SoC and monitor GPU temp in your logs.

What I sourced

2 of 12 checked

How I wired it

Solar panel1Charge controller

MC4 connectors. Keep cable runs short to minimize voltage drop.

Charge controller2SLA battery

Fused at 5A. Controller handles float charging and low-voltage cutoff.

Battery35V buck converter

Fused 12V input. The buck converter provides clean 5V 4A to the Jetson.

5V rail4Jetson Nano

Power via the barrel jack, not USB. USB cannot deliver enough current.

PIR sensor5Jetson GPIO

3.3V trigger pin wakes the Jetson from deep sleep on motion detection.

CSI camera6Jetson

Ribbon cable through a sealed cable gland. Strain relief inside the box.

INA2197Jetson I2C

Inline on the 5V rail. Logs current and voltage per inference cycle.

Power budget lab

Will it survive the night?

This is my planning model for sizing the solar panel and battery. Adjust the sliders to see how conditions affect autonomy.

Power readout

Solar harvest71 Wh/day
Daily draw36 Wh/day
Battery autonomy1.8 days
Energy surplus+97%
SD card runway999 days
Readiness88%

The software stack

Wake manager

Listens for PIR GPIO interrupt. Wakes the camera and inference pipeline, then returns to sleep after timeout.

Capture service

Grabs a still frame from the CSI camera on PIR trigger. Timestamps and buffers for inference.

Inference engine

Runs MobileNet-SSD on the Jetson GPU. Returns species label, confidence score, and bounding box.

Logger

Writes detection records to SD: timestamp, species, confidence, crop, full frame, temperature, battery voltage.

Health monitor

Tracks GPU temp, battery SOC, SD card usage, and solar input. Saves hourly summaries.

Power optimizer

Dynamic sleep intervals based on time of day. Deep sleep at night, shorter intervals at dawn/dusk.

Ethics I follow in the field

1

I deploy only on land I own or have explicit permission to monitor.

2

The camera never records audio. It captures stills, not continuous video.

3

All data stays on the local SD card. Nothing leaves the device over a network.

4

I avoid nesting sites, dens, and sensitive habitats. The observer should not change what it observes.

5

If the device disturbs wildlife behavior, I move it. Observation is the goal, not intervention.