Object detection with no identification

Hi everyone,

This will be an odd question as I am new in programming for image detection and classification. I am collecting data from an experiment using a jetson nano, a liquid going through a channel carrying objects for detection which need to be initially classify by an expert. I want to create a software to detect any potential object, creating a bounding box for objects of certain size. I have found that this is integrated into the detection models (by checking jetson inference documentation) but I have not found any tutorials on how to implement it as I am not familiar with the technical terms.

Any tutorial or tip on this topic would be really appreciated.
Thank you in advance for your time and support,
Jorge Silva

Hi @jorgersilvac, you can follow this tutorial for collecting + annotating your own data, training your own detection model, and deploying it on Jetson Nano:

Thank you for your response.
I read online and I found “Tracking Objects Using Contours” but most of the samples are done on cpu using opencv.
Does anyone know of something similar done on GPU ?

If you are talking about tracking, DeepStream does tracking on the GPU.

jetson-inference does object detection on the GPU, but not tracking.

Thank you for your time, I am a bit confused at the moment. I should explain better what I want to do, I am not quite sure how to implement it.

I am collecting data from an experiment setup using a camera and a jetson nano, a liquid going through a channel carrying objects for detection and counting.
I am not planning to do classification live, I am more concerned with accuracy rather than speed. I will use all the pixels available from the sensor as I believe this will improve accuracy of the classification, 4056 x 3040 pixels HQ PiCamera.

  1. I want to identify any potential object from the video stream and create a bounding box around the object. I read about selective search or using contours for this, I am not sure if this is the right approach.

  2. I want to extract only one image of each potential object using the bounding box, I think object tracking might be the right choice but not sure. The number of objects at any time has not exceed 5.

  3. I will classify the images to identify its class using the jetson-inference docker, move all objects of a class into a folder and count the number of objects.

I believe this summarizes what I want to do. I hope it is clear enough otherwise I will clarify any doubt you might have.
Any advice anyone could give me on how to tackle this task will be greatly appreciated.

Thank you for your time,
Jorge

Hi Jorge, I can’t really comment on the specifics of your approach, but for tracking I would recommend looking into either DeepStream or OpenCV. If you train a DNN model then you can use DeepStream for object counting.

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.