Seeking Guidance for Developing a Gender-Sensitive People Counter Using Body Detection

Hello OpenCV Community,

I am currently working on a project to develop a people counter system that not only tracks entrances and exits but also incorporates gender recognition. The primary challenge I’m facing is the need to rely on body detection for this task, as the camera setup does not consistently capture faces with sufficient quality for reliable facial recognition. The pedestrians’ full bodies are visible, which I believe could be leveraged for gender determination.

Given these constraints, I have a few questions and am seeking advice on the following points:

  1. Feasibility: Is it feasible to accurately determine a person’s gender based on body detection algorithms using OpenCV? If so, what are the limitations or challenges I should be aware of?
  2. Algorithm Recommendations: Could you recommend any specific algorithms or methods within OpenCV that are particularly suited for body-based gender recognition? Any guidance on approaching this problem would be greatly appreciated.
  3. Data and Training: Are there any publicly available datasets that could help in training a model for this purpose? Additionally, advice on how to approach the training process for gender recognition based on body characteristics would be very helpful.
  4. Performance Optimization: Considering the system needs to track both entrances/exits and gender, what are some strategies to optimize performance, especially in a live environment with multiple people moving simultaneously?

I’m relatively new to this level of customization in OpenCV and would greatly appreciate any insights, tutorials, or examples you could share that might help me navigate these challenges.

Thank you for your time and assistance!

OpenCV is not for training. It can run models though.

OpenCV does not come with models, at least not for this. bring your own model.

You need to train your own model, or use an existing trained model, or fine-tune it.

“gender” is a touchy topic in western academia. you’ll find more work outside of the west (asia), or from commercial organizations (i.e. models might not be free, but for sale or as a service).

I think AI can do as well as humans in that task. I think AI can even do a good job in the case of particular individuals that want to be unclassifiable.

what makes you think so ? do you have any paper / link / existing use-case for this idea ?

also, can you tell us, what you do, so far, to achieve those high stretched goals ?

Hello crackwitz,

Thank you for your response. I understand that OpenCV is primarily designed for running models rather than training them. I’m open to using an external model for the task and integrating it with OpenCV for the detection part. Regarding your point on gender as a touchy topic, I am aware of the sensitivities involved and my aim is to approach this project with the utmost respect for privacy and inclusivity.

I am interested in exploring AI’s potential for tasks traditionally handled by human observation and seeking to find ways in which technology can assist without overstepping ethical boundaries. If you have any suggestions for models that have been successful in similar contexts, or if you know of work particularly in non-Western academia, I would be grateful for your insights.

Best regards,

Hello berak,

Thank you for your insights. To answer your questions, my interest in the potential for using body characteristics for gender recognition comes from various studies suggesting that certain body metrics can provide cues. While I do not have a specific paper or link that directly correlates to the technology application I’m envisioning, the idea is extrapolated from existing biometric recognition research.

As for the progress of the project, we are still in the early stages, and our current focus is on defining the problem and understanding the tools and methodologies that may be employed. We have yet to select a specific model or approach for the gender recognition part. Any guidance on where to look for related use-cases or how to structure the dataset for this kind of application would be incredibly helpful.

Thank you once again, and I’m looking forward to any further advice you might have.

Best regards,