Dept. of Robotics Science and Technology,
Chubu University

People Image Analysis Human Detection

Human Detection and Action Recognition using Depth Images

Recently, depth imaging sensors such as time-of-flight (ToF) cameras or Kinect that output ranging information in real time have recently been attracting attention. We are conducting research into human detection, people-flow measurement, and action recognition as examples of object recognition from range imaging. By using ranging information, we can greatly improve the recognition performance in comparison with an ordinary visible-light camera, while utilizing previous image recognition framework.


People-Flow Measurement by Convex Shape Filtering
When human bodies are detected from depthing imaging of a ToF camera installed in the ceiling, the appearance of a human body will change greatly depending on direction, making it difficult to implement a learning base such as that described previously. We therefore focus on the convex shape of shoulder-head-shoulder of the human body and propose an approach to people-flow measurement by convex shape filtering. A Haar-like filter is disposed on both shoulders and the head, and the response value of the filter is used in rejecting convex shapes. During this time, high-speed filtering processing is enabled by the utilization of integral images. The position of the head of a human body can be detected in real time by integrating filter outputs for each of four directions. Highly accurate people-flow measurement is enabled by tracking the detected human bodies.


Action Recognition from Depth Images
We implement action recognition by extracting ranging information and time-space information from a ranging video image and using it for learning. We are working on the recognition of actions that focus on articles, from ranging imaging of a ToF camera installed in the ceiling. The system learns PSA features from a ranging video image as time-space information on the action of picking up an article and the peaks of distance histograms such as the locations of the hands as features, and constructs a multi-class classifier.


Previous Next