Vehicle Detection Model used

Autonomous Machines Jetson & Embedded Systems Jetson TX2
1

I am new to deep learning and i have just started jetson-inference. Currently i am planning to detect only vehicle in real time and i would like to ask if it would be better for me to train my own model or to use the pre-trained models(where there are many types of objects classified).

Also, would doing some pre-processing on the frames be good(ie background differencing) for the object detection model or should i just stick to the original frames? In other words, is it worth doing pre-processing or can the gpu support the object detection process at a real time rate (~25fps perhaps).

3

Hi,

We have a TrafficCamNet model in our Deepstream SDK.
You can check if the performance can meet your requirement first.
https://ngc.nvidia.com/catalog/models/nvidia:tlt_trafficcamnet

This model can reach 656 fps on Xavier so it’s expected that it can meet 25 fps on TX2.
Thanks.

4

do i need to configure the model? Or can i simply plug it in like how i do in detectnet for jetson-inference? I want to try it without the Transfer Learning first if possible

5

Currently i have tried the models like ssd, googlenet but they are not very good at capturing moving objects in traffic flow. The accuracy isnt that good

6

Hi,

Have you maximized the device performance first?

$ sudo nvpmodel -m 0
$ sudo jetson_clocks

Thanks.

7

I think you misunderstood, the fps is around for the non-custom trained model(90 classes), the fps is around14. I believe if i custom trained it to only detect the classes i require(~5), the throughput should go up. My main concern is with false detection

8

Hi,

The model can be inferenced with deepstream sample directly:

And the accuracy is roughly 83% which mentioned our TLT page:

It’s recommended to give it a try to see if this can meet your requirement.
Thanks.

10

I have tested the sample and the detection rate is good on the sources/streams i am using. However, how do i deploy this model in my own program? I don’t really get the documents provided on say parameter extractions.

12

Hi,

You can use it through the deepstream SDK.
Here are some example for Deepstream for your reference:

C++ : /opt/nvidia/deepstream/deepstream-5.0/sources/apps/sample_apps/
Python : GitHub - NVIDIA-AI-IOT/deepstream_python_apps: DeepStream SDK Python bindings and sample applications

Thanks.