rtdetr-coco-r18

Real-time object detection on camera frames is achieved by processing RGB camera inputs through an ONNX runtime. This skill identifies 80 COCO-category objects per frame, emitting bounding boxes, labels, and confidence scores to enable robotic scene understanding.

To run RT-DETR object detection, you need an RGB camera input and an ONNX runtime environment. The skill processes the camera stream to publish ObjectsMetadata events containing detected bounding boxes, labels, and confidence scores.

Real-time object detection outputs data by publishing ObjectsMetadata events for each detected object. These events contain bounding boxes, labels, and confidence scores, which can be consumed for augmented reality overlays or robotic obstacle avoidance.

Yes, you can use RT-DETR on COCO categories for robotic obstacle avoidance. It detects 80 common objects in real-time camera frames, providing the spatial awareness needed to navigate around obstacles.

The best way to integrate ONNX object detection is to deploy this skill on your robot's camera tee. It handles real-time frame processing and emits standardized ObjectsMetadata events for immediate downstream consumption.

Limitations of using ONNX runtime for real-time object detection include the strict requirement for an RGB camera input. Detection is also constrained to 80 COCO categories, meaning custom objects outside this set cannot be identified.