POST-CREST Project Gruup

  We are developing a camera system which automatically detects and tracks people in order to understand the scene. As for the camera, we are using the stereo camera which enables us to measure the object easily in three dimensional environments.
  Our final goal is to understand what is going on in the city through the images and three- dimensional information by applying the techniques such as detection of suspicious person, detection of the flow of the people and so on. Recently, from deep-Network, our research is about to further evolve.

Human Detection with Subtraction Stereo

  With our subtraction stereo, it is also able to detect human and get depth outdoors. And more, regarding stereo measurement that tends to be unstable, measurement can be limited to the foreground extracted by differential stereo, enabling robust measurement and high speed measurement.
img01
  Algorithm of subtraction stereo
img02
  Flow of people detection

-Related papers-

Fast and Stable Human Detection Using Subtraction Stereo with HOG Features

  It is important to detect a human from a camera image automatically for the measurement of human flow from a surveillance camera. We have proposed a stereo camera called "Subtraction Stereo" which focuses distant measurement on foreground regions. We implement a fast and stable human detection using the subtraction stereo with Histograms of Oriented Gradients (HOG) features.
JointHOG
  Person detection using JointHOG feature amount



Object tracking system

  In order to understand human actions or obtain flow of people in the scene, it is necessary to implement the tracking system and measure people sequentially. We are proposing a tracking system using Kalman filter. This Kalman filter predicts the state of people for next frame applying constant-velocity model as its transition model, and the tracking is realized associating this predicted states with measured data.
Human_Tracking
  Human tracking

-Related papers-

Subtraction disparity image

  There is a characteristic that background difference by color image is weak against light change. So, we propose a method as parallax image difference by using the image which is influenced by illumination change.
  Detect the person region in the order "Get background parallax image" -> "Difference with entered parallax image" -> "Determine whether it is a person based on the distance value and the size of the foreground area".

Person identification using deep learning

  A person is detected from the photographed image, and a face is detected from the person candidate region. Using the detected face as an input, the degree of similarity to each person is outputted by a model learned beforehand. When the degree of similarity for each person is equal to or less than the threshold value, it is identified as a suspicious individual. On the other hand, when the threshold value is exceeded, a person showing the highest degree of similarity is set as the identification result. When face detection can not be performed, person identification judgment is interpolated by tracking a person candidate region.

Measurement of pedestrians in a city environment using the stereo vision system

  We are developing a camera system which automatically detects and tracks people in order to understand the scene. As for the camera, we are using the stereo camera which enables us to measure the object easily in three dimensional environments. Our final goal is to understand what is going on in the city through the images and three- dimensional information by applying the techniques such as detection of suspicious person, detection of the flow of the people and so on.

  • CREST Program Download
      Documentation of how to install and use the program can be found in CrestProgram.pdf.
      DLL file is provided to use the Crest software within other programs. A header file is given to simplify the development.
      A video shows how the Crest program performs. The movie represents the elaboration of real-time captured images. Green pixels are labeled like shadow, blue pixels are the points detected like objects or pedestrian, and then analyzed.
  • CrestProgram Download(31.5MB)
    CRESTPV
      Download the promotion video of CREST Project (however, it is Japanese)
    for Windows( .wmv 76MB )
    for Mac ( .mp4 108MB )
    CRESTJPG

    -Related papers-

    Object tracking system

      In order to understand human actions or obtain flow of people in the scene, it is necessary to implement the tracking system and measure people sequentially. We are proposing a tracking system using Kalman filter. This Kalman filter predicts the state of people for next frame applying constant-velocity model as its transition model, and the tracking is realized associating this predicted states with measured data.
    Tracking_Example
      Tracking Result

    -Related papers-

    Shadow detection / Object recognition / Application of the System

      We present a visual scene description and interaction framework for pedestrian and mobile objects detection and tracking applications. The framework is built upon a previously developed stereo vision system; detect and ignore the shadows in an image, label, and track, and classify objects in the scene. The proposed algorithms raise up the information level in order to allow to query about the scene using natural language or semantic operators and give a simpler interface with other technologies.
    shadow_detection
      Examples of shadow detection
    (Blue: detected object region, Green : shadow region)

    ClassificationAndTracking
      Examples of object recognition and trajectories of objects

    -Related papers-

    Fast and Stable Human Detection Using SubtractionStereo with HOG Features

      It is important to detect a human from a camera image automatically for the measurement of human flow from a surveillance camera. We have proposed a stereo camera called "Subtraction Stereo" which focuses distant measurement on foreground regions. We implement a fast and stable human detection using the subtraction stereo with Histograms of Oriented Gradients (HOG) features.
    Human Detection Example
      Result of human detection

    -Related papers-

    Multi-Object Segmentation in a Projection Plane

      We propose a method for segmentation of multiple humans in a projection plane. The main contribution is how the image sequences that include partial occlusion of the foreground objects can be accurately segmented using mean shift clustering in real-time processing. The proposed method is suitable for inside a medium-sized environment, such as a room and an entrance.
    Segmentation
      Multi-Object Segmentation