Contents

Object Tracking

Object tracking is supported by DLStreamer through gvatrack element. It is typically inserted into a video pipeline after gvadetect element.

There are several tracking types supported, for example, short-term-imageless, zero-term, zero-term-imageless. By default, zero-term tracking is enabled which requires object detection to be run on every frame.

More details object tracking, tracking types, performance considerations, examples can be found here.

Usage

Note: By default, tracking is disabled.

To enable object tracking in a video pipeline that has a UDFLoader element, two elements gvapython and gvatrack need to be inserted after the udfloader element in the pipeline. When tracking is enabled, the metadata will include the object ids starting from sequentially from 1 for each detection.

  1. gvapython element adds detected ROIs to video frame. This is required by the tracking element to assign object ids.

  2. gvatrack adds tracking info.

To disable tracking, the above elements (if present) following the udfloader element need to be removed from the pipeline. When tracking is disabled, the metadata will include the object id as None for each detection.

Here is an example for enabling tracking for person detection pipeline that uses Geti UDF. "pipeline": "multifilesrc loop=TRUE stop-index=0 location=/home/pipeline-server/resources/classroom.avi name=source ! h264parse ! decodebin ! queue max-size-buffers=10 ! videoconvert ! video/x-raw,format=BGR ! udfloader name=udfloader ! gvapython class=AddDetectionRoi function=process module=/home/pipeline-server/gvapython/detection/add_roi.py name=add_roi ! gvatrack tracking-type=short-term-imageless ! appsink name=destination"

Specific metadata format is expected out of the UDF for it work with the gvapython that adds ROIs. For more details on the gvapython script, refer to gva_python_script

Recommendations

  • Tracking accuracy is reliant on model’s accuracy in detecting objects. For skipped detections, other tracking types like short-term-imageless may be explored which doesn’t expect detection on every frame.

  • When running multiple pipelines with tracking enabled, the object ids would still sequentially start from 1 for each pipeline. It can be coupled with either pipeline tag or unique published topic identifying results associated with a specific pipeline.

Known Issues

Currently, visualizer doesn’t show the object ids along with the overlayed detection/labels.