How to 'use' object detection result

Hello everyone,
I am new to programming and the world of machine learning.
I understand there is a lot of fantastic open source dataset like ResNet, GoogleNet etc, that is capable to detect everyday object like cars, person etc.
My question is how can I make use of the result - let say if I want to start a video recording based on object detected on the frame. e.g only start recording if there is only car and people detected. In a high level senses, how I hook the software together to allow my jetson nano to have the capability.
As far as my research goes, I assume I can use a python script and ‘if, else’ statement?
Sorry for the noob question but I just hope someone can give me guidance on how and where I should look more into this

Thanks a lot!


It’s recommended to check our Deepstream SDK.

You can add the function like if a bbox is detected, then encoding the frame.


1 Like

Thanks for the reply. After some reading up the documentation and course on DLI I am still confused as to how to get started about it. I understand there is a feature on testapp-5 with smart video recording. But how do I implement it, like as you described (add function like encoding the frame when bbox is detected). Can you elaborated more and describe the steps required. I cannot find any tutorial on this topic. Thank you very much


First, it’s recommended to check our basic pipeline that can detect car and people with nvinfer.


If an object is detected, the app will call osd_sink_pad_buffer_probe to draw the bounding box.
So you can add the smart record component into this (or in the same way).

As you already know, deepstream-test5-app is a sample to demonstrate how to use SR component.
You can follow it to add the SR component into the deepstream-test1.

In general, SR can be initiated with NvDsSRCreate and starts with NvDsSRStart.
For more details, please check the below document for reference: