DriveWorks SDK Reference
3.5.78 Release
For Test and Development only

Single Camera Template Tracking
SW Release Applicability: This tutorial is applicable to modules in both NVIDIA DriveWorks and NVIDIA DRIVE Software releases.

The tutorial shows the general structure of a program that uses the 2D template tracker to track templates in a single camera.

Initialize template array on both CPU and GPU

contextHandle is assumed to be previously initialized dwContextHandle_t.

dwTemplateArray_create(&templateCPU, maxFeatureCount,
dwTemplateArray_create(&templateGPU, maxFeatureCount,

Initialize template tracker parameters with default values

Modify parameters according to tracking requirements

imageProps is the dwImageProperties structure of image to be tracked

dwImage_getPixelType(&params.pxlType, imageProps.format);

Select on which processor the tracker should be performed on. Possible options are DW_PROCESSOR_TYPE_GPU, DW_PROCESSOR_TYPE_PVA_0, DW_PROCESSOR_TYPE_PVA_1. Note that PVA is only supported on NVIDIA DRIVE platforms. Moreover, if the processor is set to PVA, the image dimensions need to be provided at initialization.

params.imageWidth = 1280;
params.imageHeight = 800;

Initialize template tracker with given parameters and bind output

dwTemplateTracker_initialize(&tracker, &params, 0, contextHandle);

Create 2 image pyramids as detection/tracking input: one is current, one is previous, the 2 pyramids work as a double buffer (or ping-pong buffer). They swap each other after each detection/tracking cycle. So new frame will always be read into pyramidCurrent and the previous frame will be stored to pyramidPrevious.

dwPyramid_create only creates pyramid image and allocates memory, pyramid will be filled in dwImageFilter_computePyramid. Top level (level 0) in pyramid is always the same size as original input image.

inputImageProps.width, inputImageProps.height are the same in tracker/detector initialization

imagePxlType should be got from dwImage_getPixelType(inputImageProps.format).

contextHandle is assumed to be previously initialized dwContextHandle_t.

dwPyramid_create(&pyramidPrevious, levelCount,
inputImageProps.width, inputImageProps.height, pxlType, context);
dwPyramid_create(&pyramidCurrent, levelCount,
inputImageProps.width, inputImageProps.height, pxlType,context);

Add bounding boxes to be tracked to the tracker. BBox can be extracted from the DNN detector, or defined by the user.

// get bbox from DNN detector or user defining
uint32_t nNewBoxes = ...;
dwBox2Df* newBoxes = ...;
*featureData.featureCount = nNewBoxes;
for (uint32_t i = 0; i < nNewBoxes; i++) {
featureData.locations[i] = {newBoxes[i].x + newBoxes[i].width/2, newBoxes[i].y + newBoxes[i].height/2};
featureData.sizes[i] = {newBoxes[i].width, newBoxes[i].height}
featureData.statuses[i] = DW_FEATURE_STATUS_DETECTED;
// copy data to GPU
dwTemplateArray_copy(&templateGPU, &templateCPU, 0);

Start tracking the loop

// swap current/previous pyramid
// so new frame will always be built in pyramidCurrent
// and old pyramid will be stored in pyramidPrevious
std::swap(pyramidCurrent, pyramidPrevious);
// CODE: Read image frame code
dwImageHandle_t image = ...;
// compute pyramid based on the new read frame image
dwImageFilter_computePyramid(&pyramidCurrent, image, 0, contextHandle);
dwTemplateTracker_trackPyramid(&templateGPU, &pyramidCurrent,
&pyramidPrevious, tracker);
// copy detected results to CPU
dwTemplateArray_copy(&templateCPU, &templateGPU, 0);
// break the loop when all features are killed
if (*featureData.featureCount == 0)
// do some stuff in CPU

Finally, free previously allocated memory.


For the full implementation, refer to Template Tracker Sample.