A data association model for analysis of crowd structure

Abstract

The paper discusses a non-deterministic model for data association tasks in visual surveillance of crowds. Using detection and tracking of crowd components (i.e., individuals and groups) as baseline tools, we propose a simple algebraic framework for maintaining data association (continuity of labels assigned to crowd components) between subsequent video-frames in spite of possible disruptions and inaccuracies in tracking/detection algorithms. Formally, two alternative schemes (which, in practice, can be jointly used) are introduced, depending on whether individuals or groups can be prospectively better tracked in the current scenario. In the first scheme, only individuals are tracked, and the continuity of group labels is inferred without explicitly tracking the groups. In the second scheme, only group tracking is performed, and associations between individuals are inferred from group tracking. The associations are built upon non-deterministic estimates of memberships (individuals in groups) and estimates obtained directly from the baseline detection and tracking algorithms. The framework can incorporate any detectors and trackers (both classical or DL-based) as long as they can provide some geometric outlines (e.g., bounding boxes) of the crowd components. The formal analysis is supported by experiments in exemplary scenarios, where the framework provides meaningful performance improvements in various crowd analysis tasks.

Efficient face detection based crowd density estimation using convolutional neural networks and an improved sliding window strategy

“…Although CNNs are computationally expensive and generally require more data for training, compared with ML algorithms, they are effective at detecting the most complex features of images (Zitouni and Śluzek, 2022).…”