System Analysis of Video Stream Coherence in the Design of Hybrid Behavior Recognition Models

Ilya Ivanovich Kovalenko; Eugenia Alexandrovna Ahremenko; Alexandr Igorevich Andreev; Danila Yurievich Apushkin; Marina Alexandrovna Barulina

Ilya Ivanovich Kovalenko Perm State University http://orcid.org/0000-0003-4450-1184
Eugenia Alexandrovna Ahremenko Perm State University https://orcid.org/0000-0003-0757-064X
Alexandr Igorevich Andreev Perm State University https://orcid.org/0000-0002-3718-4830
Danila Yurievich Apushkin Perm State University https://orcid.org/0000-0003-4459-5264
Marina Alexandrovna Barulina Perm State University https://orcid.org/0000-0003-3867-648X

Keywords: System Analysis, Video Stream Coherence, Systems Engineering, Behavior Recognition, Generalization Inversion, Sensor Noise, Hybrid Models

Abstract

Designing video analytics systems under small-data conditions requires accounting not only for the target recognition task but also for the acquisition regime of the sensor. In this study, a two-stream CNN-ViT model with time-color coding is used as an experimental instrument to analyze two contrasting observation conditions: a static camera, where spatial and temporal representations remain mutually consistent, and active tracking, where camera stabilization introduces decoherence between these streams. The experiments show an inversion of generalization behavior across these regimes. In the coherent setting, the higher-capacity configuration provides the best recognition performance. In the decoherent setting, reducing model capacity by 62.5% prevents collapse of generalization and improves accuracy by 61.2 percentage points. In addition, removing a low-consistency class from the tracking domain, which reduces the dataset by 15%, further improves system performance by 6.9 percentage points. These results indicate that, under small-data conditions, the admissible capacity of the computational core should be selected with respect to sensor behavior and the consistency of the observed motion patterns rather than from task complexity alone.