Motion segmentation from clustering of sparse point features using spatially constrained mixture models

Motion is one of the strongest cues available for segmentation. While motion segmentation finds wide ranging applications in object detection, tracking, surveillance, robotics, image and video compression, scene reconstruction, video editing, and so on, it faces various challenges such as accurate motion recovery from noisy data, varying complexity of the models required to describe the computed image motion, the dynamic nature of the scene that may include a large number of independently moving objects undergoing occlusions, and the need to make high-level decisions while dealing with long image sequences. Keeping the sparse point features as the pivotal point, this thesis presents three distinct approaches that address some of the above mentioned motion segmentation challenges. The first part deals with the detection and tracking of sparse point features in image sequences. A framework is proposed where point features can be tracked jointly. Traditionally, sparse features have been tracked independently of one another. Combining the ideas from Lucas-Kanade and Horn-Schunck, this thesis presents a technique in which the estimated motion of a feature is influenced by the motion of the neighboring features. The joint feature tracking algorithm leads to an improved tracking performance over the standard Lucas-Kanade based tracking approach, especially while tracking features in untextured regions. The second part is related to motion segmentation using sparse point feature trajectories. The approach utilizes a spatially constrained mixturemodel framework and a greedy EM algorithm to group point features. In contrast to previous work, the algorithm is incremental in nature and allows for an arbitrary number of objects traveling at different relative speeds to be segmented, thus eliminating the need for an explicit initialization of the number of groups. The primary parameter used by the algorithm is the amount of evidence that must be accumulated before the features are grouped. A statistical goodness-of-fit test monitors the change in the motion parameters of a group over time in order to automatically update the reference frame. The approach works in real time and is able to segment various challenging sequences captured from still and moving cameras that contain multiple independently moving objects and motion blur. The third part of this thesis deals with the use of specialized models for motion segmentation. The articulated human motion is chosen as a representative example that requires a complex model to be accurately described. A motion-based approach for segmentation, tracking, and pose estimation of articulated bodies is presented. The human body is represented using the trajectories of a number of sparse points. A novel motion descriptor encodes the spatial relationships of the motion vectors representing various parts of the person and can discriminate between articulated and non-articulated motions, as well as between various pose and view angles. Furthermore, a nearest neighbor search for the closest motion descriptor from the labeled training data consisting of the human gait cycle in multiple views is performed, and this distance is fed to a Hidden Markov Model defined over multiple poses and viewpoints to obtain temporally consistent pose estimates. Experimental results on various sequences of walking subjects with multiple viewpoints and scale demonstrate the effectiveness of the approach. In particular, the purely motion based approach is able to track people in night-time sequences, even when the appearance based cues are not available. Finally, an application of image segmentation is presented in the context of iris segmentation. Iris is a widely used biometric for recognition and is known to be highly accurate if the segmentation of the iris region is near perfect. Non-ideal situations arise when the iris undergoes occlusion by eyelashes or eyelids, or the overall quality of the segmented iris is affected by illumination changes, or due to out-of-plane rotation of the eye. The proposed iris segmentation approach combines the appearance and the geometry of the eye to segment iris regions from non-ideal images. The image is modeled as a Markov random field, and a graph cuts based energy minimization algorithm is applied to label the pixels either as eyelashes, pupil, iris, or background using texture and image intensity information. The iris shape is modeled as an ellipse and is used to refine the pixel based segmentation. The results indicate the effectiveness of the segmentation algorithm in handling non-ideal iris images.