The aim of this project is to develop intelligent systems for detection and tracking rigid and non-rigid objects with non-linear motion patterns. We are as a group interested in developing novel techniques than solves tracking and detection of objects in realistic scenarios with high background clutter and low signal to noise ratio. In this regard, we have developed several different techniques which are published in various articles.
(Click on the images to go to related project web page).
Contour Tracking
Tracking the complete shape of the object provides means to perform higher level perception, such as object recognition, activity recognition, etc. We have developed a sophisticated contour tracking approach that is capable of handling complete object occlusions, background clutter and changing object and background appearance.
IR Object Tracking
Today's sensors provide means to see what humans can not see. Following this, we propose a robust object tracking for tracking objects by their body heat acquired from infrared cameras. The method is capable of tracking multiple objects simultaneously, handle large sensor motion, background clutter, and non-linear object motion.
View Geometric Approach to Multiple Camera Tracking
One of the most impressive properties of humans is their ability to solve complex visual perception problems. Among many, camera handoff is one of the most important problems for visual surveillance. We have proposed a novel approach to handle camera handoff for the moving camera scenario.
Multiple Camera Tracking and Object Identification
We have teamed up with Lockheed Martin, Orlando to merge their object identification capabilities with our multiple object tracking approach.
Analysis and Recognition of Human Actions
Analyzing human motion and recognizing the action performed by humans is one of the most significant tasks in automated surveillance, human computer interaction and psychology. Developing computational models, and recognition techniques is a very active area of research. In this regard, we have developed novel approaches which address both the recognition and development of computational models. Our main focus has been the view invariant representation and recognition techniques, which has let to a series of articles cited by most of the researchers in the field.
(Click on the images to go to related project web page).
Object Based Action Representation
Human actions are spatio-temporal events and can be considered as 3D objects. In this work, we propose to extract descriptive features of this object. Th descriptive features of a spatio-temporal events are in the form of sudden changes in shape and motion. We represent these features by analyzing a 3D volume generated from tracked object silhouettes/contours. Extracted features are shown to be robust to view point changes.
Action Recognition by Curvature Analysis
We have developed view invariant action recognition in stationary cameras. Proposed approach represent actions by analyzing curvature of trajectories and is capable of describing atomic events during an action.
View Invariant Action Recognition from Moving Cameras
In the case when the cameras are moving, observed motion in the video includes not only the actor motion but also the motion of the camera. This complicates the recognition due to extracted features do not characterize the actions. In order to perform recognition in such complicated settings, we propose a novel recognition technique, which extends the epipolar geometry of stationary camera pair to moving cameras.
Face Tracking Recognition and 3D Recovery
Tracking and recognition of human faces is a very important topic for surveillance and human computer interaction. We attach this problem from various aspects including 2D and 3D approaches.
(Click on the images to go to related project web page).
Pose Recovery
We propose a novel approach to recovery the pose of the human face. Proposed method is based on spatial relations between various face parts which are related by a so called T-structure.
Facial Motion Synthesis
An anotomical model for 3D muscle contractions can be used as a strong constraint to recover and synthesize facial motion. In the proposed work, the non-linear dynamics of the model is iteratively solved by Levenberg-Marquardet minimization method.
Face Recognition
A approach the face recognition problem holistically, assuming that the complete face contributes to perception of individuals. To overcome the illumination and facial color related issues we first detect edges and use eigenspace decomposition to a edge map passed from a regularization filter order 1, which can be considered as covering edges with a membrane.
3D Face Reconstruction
Three dimensional reconstruction of faces from images can be achieved by utilizing the symmetricity of the faces. This constraints provides an invaluable information that can be used to recover arbitrary albedo patterns of symmetric objects.
