Reconstruction of fragmented trajectories of collective motion using Hadamard deep autoencoders. (November 2022)

• Record Type: Journal Article
• Authors: Gajamannage, Kelum; Park, Yonggi; Paffenroth, Randy; Jayasumana, Anura P.
• Is Part Of: Pattern recognition. Volume 131(2022)
• Publication Date: 2022-11
• Subjects: Multi-object tracking -- Collective motion -- Deep autoencoders -- Hadamard product -- Self-propelled particles; 68T05 -- 92B20 -- 82C32 -- 62M45; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2022.108891
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: Interactions in collective motion (CM) ensure their trajectory matrix is low-rank. Low-rankness of CM is the pattern learned by our Hadamard deep autoencoder (HDA). HDA incorporates an indicator matrix into the loss function using Hadamard product. HDA is trained only using the observed trajectory segments. Performance of HDA is validated against a low-rank matrix completion framework. Abstract: Learning dynamics of collectively moving agents such as fish or humans is an essential task in research. Due to phenomena such as occlusion or change of illumination, the multi-object methods tracking such dynamics may lose the tracks of the agents which may result in fragmentations of trajectories. Here, we present an extended deep autoencoder (DA) that we train only on the fully observed segments of the trajectories by defining its loss function as the Hadamard product of a binary indicator matrix with the absolute difference between the outputs and the labels. The trajectory matrix of the agents practicing collective motion is low-rank due to mutual interactions and dependencies between the agents that we utilize as the underlying pattern that our Hadamard deep autoencoder (HDA) codes during its training. The performance of this HDA is compared with that of a low-rank matrix completion scheme in the context of fragmented trajectory reconstruction.