Pedestrian Detection, Tracking and Re-Identification for Search in Visual Surveillance Data

Visual surveillance data might encompass vast data amounts. Given the amount of data the need for search and data exploration arises naturally. Various authorities such as infrastructure operators and law enforcement agencies are confronted with search needs based on a visual description and/or behavioral patterns (motion path, activity) in order to find a ”needle in a haystack of digital data”. In this paper we present a framework which allows for an efficient search in visual surveillance archives. The paper describes following core algorithmic components of the search framework: Human detection employing pedestrian-specific shape and motion cues along with occlusion modelling; Tracking of multiple interacting pedestrians using a hierarchical spatio-temporal association scheme. Finally, pedestrian re-identification is demonstrated based on appearance matching in order to recognize a given person across a network of spatially disjoint cameras. We present results for the detection, tracking and re-identification subtasks on various challenging datasets and describe the overall framework in detail.