Towards Automated Spatio-Temporal Trajectory Recovery in Wide-Area Camera Networks

Much recent research in person re-identification has focused on improving the accuracy of matching query images from one camera view to candidates from another camera view. However, in a practical scenario, real-world surveillance system operators often must continually re-identify a person of interest through multiple views in a wide-area camera network, spatially and temporally “following” the person's trajectory. This aspect is substantially different from a traditional re-id algorithm that can only tell whether two images belong to the same person. To address this gap, we present a new algorithm to automatically reconstruct the time-stamped spatial trajectory of a person of interest moving in a camera network. With this output, a surveillance system user can easily tell where in the camera network the person of interest was located at any specific time. Since existing datasets lack the kind of annotated data needed to address this problem, we present a new dataset, RPIfield, which includes extensive trajectory annotations. We then present a novel algorithm with topology-informed transition time modeling and candidate space pruning strategies that lead to efficient trajectory reconstruction. To evaluate our method, we introduce three new evaluation metrics directly informed by practical system-level considerations and conduct extensive experiments on RPIfield.