Visual Analysis in Distributed Camera Systems
Multi-camera object analysis is an important aspect of automated multi camera surveillance applications like tracking individuals in office building corridors, cars in parking lots of airports, etc. A multi-camera system provides a viewing system capable of imaging the tracked object over a period of time. Object tracking in such a scenario holds a challenge since the same object will have different pose, shape, size, and appearance in different camera views. Moreover, different views from different cameras cause varying illumination and contrast across views. We are developing novel methods to address these challenges.
Invariant Features for Object Analysis
Project Members: Apurva Gala
We are developing a framework for tracking object motion that is partially captured by multiple non-overlapping fixed cameras. The tracking is achieved by establishing correspondence between the unique identification signatures (combination of invariant features extracted from different object views) assigned to each person, based on the views captured by the multiple cameras that fall in the object’s path. The long term goal is to extract object features (keypoints) that are invariant to rigid and as much as possible, invariant to non-rigid transformations of an object. Each feature (or keypoint) has a highly distinctive descriptor (histogram of oriented gradients) associated with it and forms the similarity criteria across object views. We have used modified implementation of SIFT to extract features invariant to scale, rotation and partially invariant to pose variations and illumination changes. Earth Mover’s Distance is used to match the object across views. Our current focus is enhancing the uniqueness of the keypoints by incorporating additional local information like texture profiles. Also, more sophisticated constraints are being explored to enhance matching.
Metrology Features for Object Analysis
Project Members: Prashanth Viswanath
The vanishing line of a reference plane, and a vanishing point for a direction not parallel to the plane can be computed. This places a restriction on the scene, that it should have a minimum number of detectable parallel lines, which can be used for computing the vanishing point and the vanishing line. By knowing the reference length, the camera height can be measured using the cross ratios. By knowing the camera height, the height of any object in the scene can be computed, again by using the cross ratios. We are extending these principles to realize a probabilistic framework for object identification and tracking across distributed camera views.