Posted on 1-1-2012 by Gert-Jan van den Braak Tags: GPU, architecture, tool, Morpheus

The Morpheus project aims to build a proof of concept for an advanced, Graphics Processing Unit (GPU) based embedded vision system, enabling contactless camera sensing of humans. The purpose of contactless sensing is unobtrusive remote monitoring of people, to avoid the use of sensors fixed on the body. The sensors used nowadays introduce vulnerability during practical use of the diagnose system, stress, pain and discomfort for healthy humans or those under medical treatment.

The ultimate goal of Morpheus is to use advanced computer vision to omit the necessity of many sensors used nowadays and to improve the quality of existing traditional sensing systems. However, video information processing requires large amounts of processing. Within Morpheus, advanced parallel processing in embedded vision systems can make contactless sensing possible. The project aims to contribute as follows:

  • An embedded vision platform will be designed and produced, supporting the constraints of contactless sensing applications. The hardware will have to meet certain constraints, such as high throughput and low power. Therefore, the platform will be based on a massively parallel processor, the GPU.
  • The project will deliver a design environment used to map vision applications onto the hardware platform. This design environment brings programmability of these platforms into reach of all, even small, companies.
  • The two applications, baby monitoring at home and neonatal care, will be developed, mapped and tested onto the hardware platform.
  • Vision algorithms, specific for the sensing of vital parameters, will be designed and adjusted to map onto the embedded vision platform. This includes software sensors used for the modelling process.
  • The project will create a demonstrator for the platform running the two applications.
The 7680 CUDA core, 22TFLOP compute cluster
Posted on 29-7-2011 by Cedric Nugteren Tags: GPU, cluster

So far, we’ve assembled and installed our 16-GPU compute cluster. Some numbers:

  • 4 compute nodes
  • 1 host node
  • 16x NVIDIA GeForce GTX570 boards, 4 per node
  • 7680 CUDA cores
  • 22TFLOPs GPU compute power
  • 2.4TB/s combined GPU memory bandwidth
  • 4KW power supply

Some pictures of the cluster:

You can download all pictures here

Speed Sign Detection and Recognition by Convolutional Neural Networks
Posted on 18-5-2011 by Maurice Peemen Tags: GPU

The objective of this project is the development of a fully trainable application for speed sign detection. Instead of using a complex chain of algorithms this solution uses a more general algorithm based on Convolutional Neural Networks (CNNs).

The advantages of this approach are:

  • A very flexible algorithm that does not require much design time due to training with examples
  • Possibility to add new recognition objectives by a simple weight set update
  • Due to the parallel nature of the algorithm it is very scalable on the next generation of computing platforms

The result of the project is a speed sign recognition application that can detect multiple speed signs in a 720×1280 HD video stream real time at 35 fps.

Demo videos that show the capabilities of the application are available on YouTube:

Or with major road detection: link

More details are described in our paper, check the attachment.

Download attachment: Pdf
Embedded Vision Architecture (EVA)
Posted on 2-3-2011 by Cedric Nugteren Tags: architecture, application, tool, FPGA

The overall EVA project objective is:

  • To develop a generic Embedded Vision Architecture (EVA) based on suitable applications
  • To apply digital design tools to develop application specific versions of this generic Embedded Vision Architecture (EVA) template
  • To further develop design tools to make flexible use of FPGAs in the designs possible
  • To demonstrate this by the designing of offspring of the generic EVA for three applications:
    1. Distributed vision for component placement machines
    2. Vision in the loop for industrial ink-jet printing
    3. Cooperating tracking cameras for Augmented Reality and Automated Vehicles
  • To realize and test the offspring and successfully integrate it in existing products

People involved: Yifan He, Zhenyu Ye

More information is available on the project website.