Unsupervised model selection for view-invariant object detection in surveillance environments

We propose a novel approach for view-invariant vehicle detection in traffic surveillance videos. Instead of building a monolithic object detector that can model all possible viewpoints, we learn a large array of efficient view-specific models corresponding to different camera views (source domains). When presented with an unseen viewpoint (target domain), closely related models in the source domain are selected for detection based on a novel discriminatively trained distance metric function, which takes into account scene geometry, vehicle motion patterns, and the generalizing ability of the models. Extensive experimental evaluation on a challenging test set, consisting of images collected from fifty different surveillance cameras, demonstrates that our unsupervised approach can outperform complex methods that utilize labeled training data from the target domain, both in terms of speed as well as accuracy.