Patterns of motion in non-overlapping networks using vehicle tracking data

Abstract

We present a systematic framework to learn motion patterns based on vehicle tracking data captured by multiple non-overlapping uncalibrated cameras. We assume that the tracks from individual cameras are available. We define the key problems related to the multi-camera surveillance system and present solutions to these problems: learning the topology of the network, constructing tracking correspondences between different views, learning the activity clusters over global views and finally detecting abnormal events. First, we present a weighted cross correlation model to learn the topology of the network without solving correspondence in the first place. We use estimates of normalized color and apparent size to measure similarity of object appearance between different views. This information is used to temporally correlated observations, allowing us to infer possible links between disjoint views, and to estimate the associated transition time. Based on the learned cross correlation coefficient, the network topology can be fully recovered. Then, we present a MAP framework to match two objects along their tracks from non overlapping camera views and discuss how the learned topology can reduce the correspondence search space dramatically. We propose to learn the color transformation in [iota][alpha][beta] space to compensate for the varying illumination conditions across different views, and learn the inter-camera time transition and the shape/size transformation between different views.

(cont.) After we model the correspondence probability for observations captured by different source/sinks, we adopt a probabilistic framework to use this correspondence probability in a principled manner. Tracks are assigned by estimating the correspondences which maximize the posterior probabilities (MAP) using the Hungarian algorithm. After establishing the correspondence, we have a set of stitched trajectories, in which elements from each camera can be combined with observations in multiple subsequent cameras generated by the same object. Finally, we show how to learn the activity clusters and detect abnormal activities using the mixture of unigram model with the stitched trajectories as input. We adopt a bag - of - words presentation, and present a Bayesian probabilistic approach in which trajectories are represented by a mixture model. This model can classify trajectories into different activity clusters, and gives representations of both new trajectories and abnormal trajectories.