I need to be able to detect and track humans from all angles, especially above.

There are, obviously, quite a few well-studied models for human detection and tracking, usually as part of general-purpose object detection, but I haven't been able to find any information that explicitly works for tracking humans from above.

  • When you say "above," do you mean humans approaching from the top down? In this case you can use existing human detection in libraries and analyze the size of the object. The closer, the smaller the distance and the larger the object. – Guilherme IA Oct 29 at 17:06
  • @GuilhermeIA Yes, top down, as in only top of head and shoulders discernable. Do existing human detection libraries account for this? – T3db0t Oct 29 at 17:14
  • "from the top" like satellite images or like surveillance cameras? – Martin Thoma Oct 29 at 20:07
  • @MartinThoma From a ceiling-mounted camera anywhere from 10-35 feet above the floor. But ideally it could be mounted anywhere at any angle, so ideally detecting humans from all possible angles – T3db0t Oct 29 at 20:21

Look at all the Haar boosted classifiers that are available with OpenCV and the dedicated class CascadeClassifier to use it. Here are a list of what the classifiers have locally...

You can try frontalface. The shoulders I searched and did not find much. You may have to create a shoulder dataset and train to detect it.

https://stackoverflow.com/a/5727231/5336017

  • 1
    Wouldn't "frontalface" be face-on, or am I missing something? I need literally the tops of heads and shoulders as distinct from other objects. – T3db0t Oct 29 at 19:39
  • As I said, these are the ready-made libraries that exist. In your case, as something is perhaps specific and you have not yet on the internet, you will need to create your own head and shoulder top dataset. From a researched on tensorflow-image-classifier. You will need a huge base of images and train them. – Guilherme IA Oct 29 at 19:46
  • 1
    Sure, I get that. I would just be surprised that there wouldn't already be a dataset for this purpose used in surveillance, drones, etc.. – T3db0t Oct 29 at 20:22

There were no pre-trained models for human that were taken from all angles (which is a somewhat extreme requirement) or specifically from above. This condition may change due to the high value of such to the surveillance industry.

There are already several projects for which there is human specific pattern recognition and there have emerged some data sets to support this more specific field of objectives, some of which are based on moving rather than stationary images. This is for two reasons.

  • Multiple frames has the potential (should computing resource utilization be overcome) to produce higher accuracy and reliability in human recognition partly because of the identification of joints in movement and partly because there is simply more information available on each object that could be a human.
  • Trajectory of pedestrians is important for both law enforcement and security and for avoiding pedestrians in vehicle automation.

These are some of the projects and download sites that are related to the goal of the question, but unfortunately do not seem to provide data sets of humans from above or pre-trained networks from that visual perspective.

Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

Not the answer you're looking for? Browse other questions tagged or ask your own question.