Skip to content

Conversation

@matteius
Copy link
Contributor

Support for the Coral TPU -- tested on Raspberry Pi 5.

Note: I had to for and update update libedgetpu to get things working on modern tensorflow.

INFO:     127.0.0.1:58172 - "POST /api/v1/detect?backend=edgetpu&confidence_threshold=0.5&return_image=false HTTP/1.1" 200 OK
2025-12-13 22:58:41,725 - api.v1.endpoints.detection - INFO - Detection request: backend=edgetpu, confidence=0.5, filename=lightnvr_api_detection_24958.jpg
2025-12-13 22:58:41,725 - backends.edgetpu.backend - INFO - Attempting to load EdgeTPU as auto
2025-12-13 22:58:41,765 - backends.edgetpu.backend - INFO - EdgeTPU found and loaded successfully
2025-12-13 22:58:41,776 - backends.edgetpu.backend - INFO - EdgeTPU backend initialized: model=ssd_mobilenet_v2_coco_quant_postprocess_edgetpu.tflite, type=ssd, device=auto, input_shape=[  1 300 300   3], tflite_source=tensorflow
2025-12-13 22:58:41,888 - api.v1.endpoints.detection - INFO - Detection completed: 1 objects found in 33.8ms
2025-12-13 22:58:41,888 - api.v1.endpoints.detection - INFO - Detected objects: {'person': 1}
2025-12-13 22:58:41,888 - api.v1.endpoints.detection - INFO - Response prepared: 1 detections, 34.1ms total
image

- Add EdgeTPU backend with SSD and YOLO model support
- Support USB and PCIe/M.2 Coral devices with device selection
- Add int8 input handling for quantized models
- Add YOLO DFL decoding for EdgeTPU-compiled YOLO models
- Add download script for Coral Model Zoo models
- Register EdgeTPU backend in factory with lazy loading
- Add EdgeTPU configuration options to config.py
@matteius matteius merged commit cc816a2 into main Dec 14, 2025
@matteius matteius deleted the feature/edgetpu-backend branch December 14, 2025 04:16
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants