A survey on horizon detection algorithms for maritime video surveillance: advances and future techniques

Autor: Yassir Zardoua, Mohammed Boulaala, Abdelali Astito
Rok vydání: 2021
Předmět:
Zdroj: The Visual Computer. 39:197-217
ISSN: 1432-2315
0178-2789
DOI: 10.1007/s00371-021-02321-0
Popis: On maritime images, the horizon is a linear shape separating the sea and non-sea regions. This visual cue is essential in several sea video surveillance applications, including camera calibration, digital video stabilization, target detection and tracking, and distance estimation of detected targets. Given the nature of these applications, the horizon detection algorithm must satisfy robustness and real-time constraints. Our first aim in this paper is to provide a comprehensive review of horizon detection algorithms. After analyzing assumptions and test results reported in the horizon detection literature, we found a high trade-off between robustness and real-time performance. Thus, our second aim is to propose and describe three workable techniques to reduce this trade-off. The first technique aims to increase the robustness against contrast-degraded horizons. The non-sea region right above the horizon mainly depicts the sky, coast, ship, or combination of these classes. Thus, the second technique suggests a way to handle such class variation. While we believe that the last two techniques require relatively little computations, the third technique concerns using an alternative convolutional neural network (CNN) architecture to avoid a significant quantity of redundant computations in a previous CNN-based algorithm.
Databáze: OpenAIRE
Externí odkaz: