Jetson Anomaly Detection Stack

This repository runs the deployed YOLO segmentation and anomaly-localization pipeline on a Jetson Orin. It connects to the Raspberry Pi ROS 2 graph through rosbridge; it does not join the robot DDS network.

Runtime

Raspberry Pi camera, depth, scan, map and pose topics
                        |
                  rosbridge :9090
                        |
                   Jetson Orin
                        |
       YOLO segmentation, tracking and localization
                        |
       images, JSONL events and visualization topics

The Compose stack starts:

• jetson_anomaly, using filipsijak2/jetson_anomaly:jetson_anomaly-dev-local
• jetson_log_viewer, exposing read-only artifact directories on port 8081

Active configuration

The application runtime configuration consists of two files:

• config/anomaly_rosbridge.yaml contains topic names, model settings, thresholds, tracking, privacy output, 3D detections and inspection settings.
• config/containers/jetson_anomaly.env contains the runtime environment loaded by Compose.

Environment values override equivalent YAML values. Keep both files synchronized when changing a duplicated setting.

The active model is yolov8n-seg.pt, the anomaly class is bottle, tracking uses bytetrack.yaml, and GPU inference uses device 0 with half precision.

The active Raspberry Pi endpoint is:

ROSBRIDGE_URL=ws://raspberry.local:9090

The configured inputs are:

• /camera/realsense/color/image_raw/compressed
• /camera/realsense/aligned_depth_to_color/image_raw/compressedDepth
• /camera/realsense/color/camera_info
• /map
• /robot_pose_map
• /scan

The configured outputs are:

• /anomaly/events
• /anomaly/events/readable
• /anomaly/markers
• /anomaly/detections_3d
• /anomaly/debug_image/compressed
• /anomaly/privacy_image/compressed
• /anomaly/map_snapshot/compressed
• /anomaly/inspection/*

Storage

Compose mounts:

• ./models at /workspace/models
• ./anomaly_logs at /home/jetson/anomaly_logs
• ./logs at /workspace/logs

Model binaries, generated images and logs are ignored by Git. See models/README.md for model placement.

Start

docker compose pull
docker compose up -d
docker compose logs -f jetson_anomaly

For a local image build:

docker compose -f docker-compose.yaml -f docker-compose.build.yaml build

Validation

python -m pip install -r requirements-yolo.txt -r requirements-rosbridge.txt
pytest -q
python -m compileall src/jetson_anomaly_detector/jetson_anomaly_detector
bash -n scripts/start_jetson_anomaly.sh

Deployment workflow

The GitHub Actions deployment job requires these repository secrets:

• TAILSCALE_AUTHKEY
• JETSON_SSH_USER
• JETSON_SSH_PRIVATE_KEY_B64 or JETSON_SSH_PRIVATE_KEY

It also requires JETSON_HOST, preferably as a repository secret. A repository variable with the same name is supported as a fallback. JETSON_STACK_DIR accepts either form and defaults to ~/jetson-stack.