Edge Impulse now supports object tracking as a post-processing feature, enabling users to turn raw bounding boxes or centroids into stable, persistent object identities across video frames. This unlocks new capabilities for counting, monitoring, and analyzing objects in motion- directly on your embedded devices.
Smarter Object Detection with Tracking
Object detection models are great at finding objects in each frame, but they don’t know if the same object appears in multiple frames. This can lead to jittery results and makes it hard to count or monitor objects over time. With object tracking, you can now follow objects as they move, reduce detection noise, and unlock new use cases like counting items on a conveyor belt or tracking vehicles on a road. For this initial release the Object Tracking algorithm is best suited to linear motion such as vehicles on a road or objects on a production line. For more technical information about the limitations and benefits of this new feature head to the docs page.
How Object Tracking Works in Edge Impulse
Stable Traces Across Frames
Our object tracking layer maintains ongoing traces of detected objects by aligning new detections with existing traces. This means you get smooth, reliable object IDs even when detections are missed for a frame or two.
Flexible for Different Models
We support both centroid-based models (like FOMO) and bounding box models (like MobileNetV2 or YOLO architectures). The system uses either the distance between centroids or the overlap (IoU) between bounding boxes to match objects across frames.
Easy to Configure and Deploy
You can enable object tracking and tune your parameters with just a few clicks in your Edge Impulse project. Once you've trained a model, upload a test video to the new Post-Processing area in Data Acquisition. Then head to the Post-processing area to configure your parameters for stable tracking. If enabled, object tracking is automatically built into your deployment to any supported target including—Linux EIM, C++ library and more.
Unlocking New Use Cases
With object tracking, you could take the new uniquely identified objects and:
- Count objects as they cross a region (e.g., products on a conveyor)
- Monitor vehicles or people over time
- Reduce false positives from jittery detections with the smoothed results
Start Tracking Today
Object tracking brings a new level of intelligence to your ML projects in Edge Impulse. It’s easy to get started, and the benefits are immediate for anyone working with video or image streams.
Sign up for your free developer account here and start building smarter, more reliable edge AI solutions today.
For more details, check out our object tracking documentation.