Calibration-free grid-based pedestrian tracking with shortcut edge for video surveillance

Multi-object tracking (MOT) has drawn considerable interest owing to its many applications, such as analyzing pedestrian and obstacle behavior for autonomous driving perception. Conventional grid-based tracking methods are typically vulnerable to detection failures, since connections between temporally adjacent frames cannot withstand such transitory errors, whereas long-term connections act as shortcuts that bridge these failure gaps. We reformulate the grid-based approach by introducing a deep-metric long-term shortcut edge for multi-object tracking, in which a Siamese CNN is used to generate similarity scores Most of the existing grid-based methods have to perform calibration before tracking, we argue that it is not necessary. Calibration tremendously restricts tracking applications and is not suited for tracking in some practical cases of human behavior analysis. Our approach, in contrast, employs a detection-based strategy to estimate the grid model mapping. As a consequence, the users can avoid the inconvenient calibration process. We validate both the long-term shortcut edge and the calibration-free strategy through experiments on several widely used benchmarks. The long-term shortcut edge raises MOTA by close to 16%, while the calibration-free strategy could produce comparable results. Comparisons against other popular methods confirm the performance benefits of our tracker.