Grega Pušnik (2012) Human detection and tracking with mobile platform and RGBD camera. EngD thesis.
For a long time detection and tracking of people has been one of the main research topics in computer vision. This is one of the key problems for efficient interaction of an autonomous robot with people. It is a complex problem due to various human poses, lighting, background complexity and other variables. In the past, researchers were mostly using 2D RGB cameras. With the arrival of a low cost RGBD camera Kinect, researchers have increasingly started using not only 2D information, but also depth information with a point cloud. In the thesis we address the problem of a human detection and tracking on a mobile platform. System is based on ROS - Robot Operating System, mobile robot IRobot Roomba and RGBD sensor Kinect. For human detection we first use the depth information of the Kinect and HOG algorithm for the initial classification. For re-detection, algorithm narrows search window and then for classification, resorts to more robust online Adaboost algorithm with Haar features. For cases where we do not positively classify or lose the human we use predictions of the Kalman filter. For robot navigation we used ROS navigation stack. That is how we implemented the whole detection-tracking system for detecting, tracking and following people.
Actions (login required)