How Much to Learn? An Information-Sufficiency Criterion for Detecting Motion Rules in Scenario Surveillance

An automated learning process is a powerful tool that extracts useful data structures by measuring information, expressing conditions, and regulating invariant tolerance. Surveillance video systems and intelligent video analysis have become increasingly valuable, as it is often challenging to discern and represent the primary motion patterns necessary for automatic analysis. Combining automated learning with video analysis enables the development of intelligent analytic systems that can operate effectively in uncertain scenarios, automatically identifying dominant motion dynamics. The ability to recognize these motion dynamics relies on the theoretical framework used to represent them and the learning process employed to identify patterns. In the literature, a state-system-based approach offers a way to describe temporal and spatial interrelationships, which is essential for understanding motion dynamics. A key aspect of this approach is determining when the number of positive learned states from a given information source becomes sufficient to detect dominant motion in a surveillance video system. This determination is crucial as it impacts the variability of movements that the monitored subjects can exhibit while adhering to the camera’s view projection and the resources required for implementation. While previous research has produced practical approaches, most have been limited to specific scenarios, underscoring the need for further investigation. In this study, we outline several advantages of establishing a formal criterion to determine when a symbolic system in a surveillance scenario has developed a robust and practical model to explain the observed motion dynamics. Our proposal is sustained by the hypothesis that a correct model can reliably account for most motion dynamics over time in an automatic learning process. The validation is performed with different real scenarios, where the central dynamic becomes learned, labeling those dynamics that the learned model cannot explain. In a real application, the approach was used to model traffic vehicles on the avenues of Queretaro City.