High velocity object detection library RRS feed

  • Question

  • I have a use case where I need to detect (just) one person and extract their estimated origin (i.e. I do not need joint information).  I need to detect as soon as possible and in real time.  The Kinect SDK seems to have a large delay if a person enters the scene with high velocity.

    For instance, if I jump from outside the range of the Kinect to in the middle of the Kinect, there is a 1 to 2 second delay before the skeleton is identified.  This is unfortunately an unacceptable delay for my application.  I've also noticed that if I enter scene in a non ideal pose (i.e. not facing the camera), the tracking does not kick in at all until I turn into a "nice" position to bootstrap the tracking algorithm.

    Does anyone have a suggestion for a library (or relevant papers) that accepts point cloud data, and can identify people in real time but with no latency for initial detection?  The only thing moving in my application's scene will be the person, so solutions that assume a static background are acceptable.

    To sum up, I'm looking for a library that can identify and track a new object introduced into a static scene with minimal latency and supporting high velocities, and output a rough bounding box in real time for processing.

    Thanks in advance


    Sunday, December 11, 2011 2:33 AM

All replies

  • I am having a very simlar problem with my application. If I get any helpful information, I will inform you.
    Friday, December 16, 2011 1:49 PM