Want to know more about Krejov hybrid approach to human pose estimation - recent paper; also want to compare specs of sensor options
Hello,
I would like to get the latest paper(s) by Dr Krejov, on his hybrid approach to human pose estimation. His 2016 paper seems to combine models with decision forests, while his recent youtube presentation brings up a small convolutional neural network. I am very intrigued by the combination of a model-based approach and a small CNN, and would appreciate a copy of the paper and hopefully access to the source code as well.
I don't have any RealSense sensors yet, but I plan on purchasing some to configure a Smart Room that can determine fall events in geriatric subjects, in conjunction with NSF funding if successful. I would also appreciate some specs and advice on how to choose these sensors. I would think a laser depth-sensor performs more quickly and robustly than a passive stereo sensor, but perhaps you may correct that perception.
Best wishes,
Michel Audette, Ph.D.
Associate Professor, Department of Computational Modeling and Simulation Engineering,
Graduate Program Director, Biomedical Engineering Institute,
Old Dominion University,
Norfolk, VA.
-
Hi @Maudette The most recent paper credited to Phillip Krejov that I could find is a 2017 one titled 'Guided optimisation through classification and regression for hand pose estimation'.
https://www.sciencedirect.com/science/article/abs/pii/S107731421630193X
The official RealSense blog also recently published an article on using depth camera technology for health monitoring, including fall risk.
https://www.intelrealsense.com/virtual-caregiver/
Intel has a 'which device is right for you' guide that compares the features of the models in the RealSense product range.
https://www.intelrealsense.com/which-device-is-right-for-you/
Full technical data sheet documents for the 400 Series stereo depth cameras and L515 lidar depth camera can be accessed at the link below:
https://dev.intelrealsense.com/docs/datasheets
The L515 is best suited to controlled lighting conditions, and sunlight that is less than 500 lux in strength. If you plan to monitor 24 hours a day including in night-time conditions with the lights off and in highly variable natural lighting conditions, a 400 Series depth camera will likely be the best choice for your needs. If you are monitoring from a distance, the new RealSense D455 model will provide the best depth measuring accuracy over distance.
-
Hi Marty,
the 2017 paper is similar to the other one that I saw, in that the pose estimation combines Randomised Decision Forest with tubular models, rather than CNNs with models as described in that Youtube of yours. Can you ask him about details of the CNN, or ask him to get in touch with me? I tried contacting him in the past (a few weeks ago), and most recently by Twitter yesterday, but I have not heard back.
I would like to find out also how to access the source code if possible (is it in Skeleton Tracking SDK?).
Otherwise, I don't see what is the point of these Youtubes, if not to potentiate the use of Intel sensors through state-of-the-art tracking algos.
Thanks for your kind support.
Michel
-
I am unable to arrange for Phillip Krejov to talk to you or contact him myself on your behalf. Intel staff typically cannot reply to email enquiries from the public. Questions must go through a support forum such as this one or the RealSense GitHub (link below).
https://github.com/IntelRealSense/librealsense/issues/
Even then, questions will usually be responded to by Intel members who are authorized to do so, such as members of the RealSense support team (of which I am one).
Intel's official RealSense YouTube videos are typically not designed to provide a complete solution but instead provide information that acts as a starting point for RealSense users' own project ideas. Pre-made example tutorials with source code are provided at the link below though:
https://github.com/IntelRealSense/librealsense/tree/master/examples
-
Hi Marty,
I'm writing a proposal and actually citing this forum. I'm intrigued by the statement
If you plan to monitor 24 hours a day including in night-time conditions with the lights off and in highly variable natural lighting conditions, a 400 Series depth camera will likely be the best choice for your needs.
Why would a stereo camera perform better than the lidar camera L515 in darkness? It seems to me, in complete darkness, a passive stereo algorithm has no light to reflect at all, and therefore no signal, whereas the lidar is supplying its own laser light.
Right now, I am leaning towards either a Lidar configuration or a hybrid configuration with both Lidar and some complement such as a 400 Series.
Cheers,
Michel
PS: looking at LinkedIn, I can see why Philip Krejov is difficult to reach; he's now in LA working on photorealism. Pity that he didn't write it up before leaving and no one picked it up.
-
RealSense 400 Series stereo technology is categorized as Active Stereo, not passive.
https://www.intelrealsense.com/stereo-depth/
The L515 is a camera that depends on being indoors or in controlled lighting and can also have its IR interfered with by sunlight. The FAQ on the official L515 information page says: "Infrared light from the sun can interfere with the performance of the device which can degrade the quality of the depth images when used outdoors. As a power efficient LiDAR camera, L515 will perform best indoors or in controlled lighting conditions".
The 400 Series cameras have a Vision Processor D4 circuit board that can make automatic adjustments for real-time lighting conditions. The IR imager component can operate in low-light conditions, and in darkness it can take advantage of any light sources available. This could include the camera's IR Emitter component, or an external light source such as an external projector or an IR illuminator.
https://github.com/IntelRealSense/librealsense/issues/2000
In the absence of ambient light, the 400 Series camera can cast a semi-random dot pattern from its projector onto the objects in a scene. It can then use those dots as a "texture source" to analyse the dots for depth information.
Because the IR imager on the 400 Series does not have an IR-Cut filter, it can see all visible light frequencies. The D435 model may be able to see better in the dark than the D415 model as it has a wider IR imager that can let in more light, though this can also create greater noise when the light level increases.
Please sign in to leave a comment.
Comments
5 comments