Good detection poor tracking

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1

jetson xaxier AGX
deepstream4
jetpack4.3

Hi,

I’ve created and a custom model with which I get very good detections 95-99% accuracy

however when i run a video the tracker id changes a lot and sometimes I get 2 detection boxes around the same item with 2 different tracker ids.

Do I need to create a custom tracker? I am currently using the built in libnvds_nvdcf.so

Are there any other parameters that can be added to improve tracking?

Is there a better tracker that can be used or is this the limit with DS4?

thanks

%YAML:1.0

NvDCF:
maxTargetsPerStream: 30 #
filterLr: 0.11 #
gaussianSignma: 0.75 #

minDetectorConfidence: -1 #
minTrackerConfidence: 0.6 #

featureImgSizeLevel: 5 #
SearchRegionPaddingScale: 3 #

maxShadowTrackingAge: 27 #
probationAge: 12 #
earlyTerminationAge: 2 #

minVisibiilty4Tracking: 0.2 #

2

is there a newer version of libnvds_nvdcs.so that will work with ds4?

with more parameters?

I am currently using a version dated 14th December 2019

thanks

4

No, I can only suggest to move on the latest version, we just release the DeepStream SDK 5.1, see https://developer.nvidia.com/deepstream-getting-started