Crowd flow estimation from calibrated cameras
- PDF / 10,202,564 Bytes
- 12 Pages / 595.276 x 790.866 pts Page_size
- 49 Downloads / 254 Views
ORIGINAL PAPER
Crowd flow estimation from calibrated cameras Igor Almeida1
· Claudio Jung1
Received: 6 September 2017 / Revised: 13 July 2020 / Accepted: 29 September 2020 © Springer-Verlag GmbH Germany, part of Springer Nature 2020
Abstract Many crowd analysis methods rely on optical flow techniques to estimate the main moving directions. In this work, we propose a crowd flow filtering approach for calibrated cameras that can be coupled to any generic optical flow method. It projects the input optical flow to the world coordinate system, performs a local motion analysis exploring a Social Forces Model and then projects the filtered flow back onto the image plane. The method was tested on publicly available datasets involving crowded scenarios used in conjunction with different optical flow methods, and results indicate that the proposed filtering method provides coherent crowd flows when coupled to the tested methods. Keywords Crowd flow · Optical flow · Crowd analysis
1 Introduction Crowds may arise in busy streets, sporting events, music concerts, protests, among others. Tracking individuals in a dense crowd is a challenging task, and many crowd analysis techniques [9,26,35] use the optical flow as an estimate of the crowd motion. Despite the existence of many optical flow techniques (see [13] for a survey), they are mostly generic-purpose methods (i.e., they try to find local correspondences in generic scenarios). However, people in a crowd typically move in an orderly manner, and neighbors in a crowd usually have similar movement patterns (speed and orientation). As a consequence, generic-purpose optical flow methods might not be adequate for crowd tracking. In fact, most methods are prone to flow errors due to sensor noise and color/texture ambiguity, which might be incompatible with the expected motion of a real crowd. Even very accurate methods (in the limit, the ground-truth flow) might capture information that is not relevant for crowd analysis, such as individual arms and legs motion. In this paper, we propose a fast post-processing method for obtaining coherent crowd flows that explores the expected behavior of real crowds in scenes captured by a calibrated
B
Igor Almeida [email protected] Claudio Jung [email protected]
1
Federal University of Rio Grande do Sul, Porto Alegre, Brazil
static camera. It takes as input an initial flow (obtained by any optical flow method) and explores the camera parameters to estimate coherent neighborhoods in the world coordinate system, which are used to smooth the initial flow based on psychosocial aspects of crowded scenes. Then, the filtered flow is projected back onto the image plane. The remainder of this paper is organized as follows: Sect. 2 reviews the state of the art in optical flow estimation, in particular those used by crowd analysis techniques. The proposed method is introduced in Sect. 3, and Sect. 4 shows the experimental results. Finally, conclusions are presented in Sect. 5.
2 Related work Correctly estimating the crowd flow of a scene
Data Loading...
