Object Tracking#

Object tracking types#

gvatrack is an object tracking element, typically inserted into a video analytics pipeline right after the object detection element gvadetect and can work in the following modes as specified by the tracking-type property:

tracking-type	Max detection interval (inference-interval property in gvadetect)	Algorithm uses detected coordinates and trajectory extrapolation	Algorithm uses image data	Notes
short-term-imageless	<= 5	Yes	No	Assigns a unique id to objects and generates object position for frames on which object detection was skipped. Fast algorithm that extrapolates object trajectory from previous frame(s) without access to image data.
zero-term	1 (every frame)	Yes	Yes	Assigns a unique id to objects, and requires object detection run on every frame. Takes into account object trajectory as well as a color histogram of object image data.
zero-term-imageless	1 (every frame)	Yes	No	Assigns a unique id to objects, and requires object detection run on every frame. Fastest algorithm as based on comparing object coordinates on current frame with objects trajectory on previous frames.
deep-sort	1 (every frame)	Yes	Yes	Assigns a unique id to objects using Kalman filter for motion prediction and deep learning features for re-identification. Robust algorithm that minimizes ID switches by combining appearance and motion cues. Requires feature extraction model (e.g., mars-small128).

Additional configuration#

Additional configuration parameters for object tracker can be passed via the config property of gvatrack. The config property accepts a comma separated list of KEY=VALUE parameters. The supported parameters are described below:

Tracking per class#

Configurable via tracking_per_class parameter. It specifies whether the class label is considered for updating the object_id of an object or not. When set to true, a new tracking ID will be assigned to an object when the class label changes. When set to false, the tracking ID will be retained, based on the position of the bounding box, even if the class label of the object changes due to model inaccuracy. The default value is true.

Example:

... gvatrack config=tracking_per_class=false ...

Maximum number of objects#

Configurable via max_num_objects parameter. It specifies the maximum number of objects that the object tracker will track. On devices with less computing power, tracking a smaller number of objects can reduce compute and increase throughput.

Example:

... gvatrack config=max_num_objects=20 ...

Sample#

Refer to the vehicle_pedestrian_tracking sample for a pipeline with gvadetect, gvatrack, and gvaclassify elements.

Deep SORT Tracking#

Deep SORT (Simple Online and Realtime Tracking with a Deep Association Metric) is an advanced tracking algorithm that combines:

Motion prediction: Kalman filter for predicting object position based on velocity and position history
Appearance features: Deep learning re-identification model that extracts 128-dimensional feature vectors to distinguish objects
Robust association: Combines IoU (Intersection over Union) and cosine distance metrics to match detections with existing tracks

Feature Extraction Model#

Deep SORT requires a feature extraction model that generates 128-dimensional feature vectors for person re-identification. The recommended model is mars-small128, which can be downloaded using:

./samples/download_public_models.sh --model-name mars-small128

This downloads both FP32 and INT8 quantized versions of the model.

Usage Modes#

Deep SORT supports two modes of operation:

(1) Internal Feature Extractor#

The gvatrack element performs feature extraction internally. Inference runs on CPU by default.

gvatrack tracking-type=deep-sort feature-model=/path/to/mars-small128/FP32/mars-small128.xml

Advantages: Simpler pipeline, automatic feature extraction per detection

(2) External Feature Extractor#

Use gvainference before gvatrack to perform feature extraction. This allows running inference on GPU or other devices for better performance.

gvainference model=/path/to/mars-small128/FP32/mars-small128.xml device=GPU inference-region=roi-list ! \
queue ! \
gvatrack tracking-type=deep-sort

Advantages: Device flexibility (CPU/GPU/NPU), potentially higher throughput

Configuration Parameters#

Deep SORT behavior can be fine-tuned using the deepsort-trck-cfg property with comma-separated KEY=VALUE parameters:

Available Parameters#

Parameter	Default	Description	Tuning Guidelines
`max_iou_distance`	0.7	Maximum IoU(Intersection over Union) distance threshold for matching detections to existing tracks based on bounding box overlap	Lower values (0.5-0.6) = stricter spatial matching, less tolerance for movement. Higher values (0.7-0.8) = more lenient matching, better track continuity but higher risk of identity switches
`max_age`	30	Maximum number of frames a track survives without detection before deletion	Increase to 60-90 to reduce object loss during occlusions or missed detections. Lower values delete tracks faster, good for crowded scenes
`n_init`	3	Number of consecutive detections required to confirm a new track	Lower values (1-2) = faster initialization but more false positives. Higher values = more reliable but slower confirmation
`max_cosine_distance`	0.2	Maximum cosine distance threshold for appearance feature matching between detections and tracks	Increase to 0.3-0.4 to handle lighting changes, better for similar-looking objects, viewing angles, or appearance variations. Lower values = stricter appearance matching
`nn_budget`	100	Maximum number of appearance features stored per track	Higher values = better re-identification, good for extended tracking scenarios but more memory. Typical range: 50-150

Examples for Common Tuning Scenarios#

Reducing object loss (tracks disappearing too quickly):

gvatrack tracking-type=deep-sort deepsort-trck-cfg="max_age=60,max_cosine_distance=0.3"

Handling fast-moving objects:

gvatrack tracking-type=deep-sort deepsort-trck-cfg="max_iou_distance=0.6,max_age=45"

Crowded scenes with occlusions:

gvatrack tracking-type=deep-sort deepsort-trck-cfg="max_age=90,max_cosine_distance=0.35,nn_budget=150"

Conservative tracking (minimize false positives):

gvatrack tracking-type=deep-sort deepsort-trck-cfg="n_init=5,max_cosine_distance=0.15"

Example Pipeline#

gst-launch-1.0 filesrc location=video.mp4 ! decodebin ! \
  gvadetect model=person-detection.xml ! \
  gvainference model=mars-small128.xml device=GPU inference-region=roi-list ! \
  gvatrack tracking-type=deep-sort \
    deepsort-trck-cfg="max_age=60,max_cosine_distance=0.3" ! \
  gvawatermark ! videoconvert ! autovideosink

How to read object unique id#

The following code example iterates all objects detected or tracked on the current frame and prints object unique id and bounding box coordinates.

#include "video_frame.h"

void PrintObjects(GstBuffer *buffer) {
    GVA::VideoFrame video_frame(buffer);
    std::vector<GVA::RegionOfInterest> regions = video_frame.regions();
    for (GVA::RegionOfInterest &roi : regions) { // iterate objects
        int object_id = roi.object_id(); // get unique object id
        auto bbox = roi.rect(); // get bounding box information
        std::cout << "Object id=" << object_id << ", bounding box: " << bbox.x << "," << bbox.y << "," << bbox.w << "," << bbox.h << "," << std::endl;
    }
}

Performance considerations#

Object tracking can help improve performance of both object detection (gvadetect) and object classification (gvaclassify`) elements

Object detection: short-term-imageless tracking types enable reducing object detection frequency by setting the inference-interval property in the gvadetect element.
Object classification: if an object was classified by gvaclassify on frame N, you can skip classification of the same object for several next frames N+1,N+2,… and reuse last classification result from frame N. Reclassification interval is controlled by the reclassify-interval property in the gvaclassify element.

See the sample pipeline below:

gst-launch-1.0 \
... ! \
decodebin3 ! \
gvadetect model=$DETECTION_MODEL inference-interval=10 ! \
gvatrack tracking-type=short-term-imageless ! \
gvaclassify model=$AGE_GENDER_MODEL reclassify-interval=30 ! \
gvaclassify model=$EMOTION_MODEL reclassify-interval=15 ! \
gvaclassify model=$LANDMARKS_MODEL ! \
...

It detects faces every 10th frame and tracks faces position for the next 9 frames. The age and gender classification is updated every second, the emotion classification is updated twice a second, and the landmark points are updated every frame.