Deep learning-based fishermen's behaviour recognition with wearable devices to enhance maritime safety

1 de fevereiro de 2026

From the perspective of fishermen surveillance, this study investigates fishermen’s behaviour and operational patterns to enhance workplace safety during fishing activities. This paper introduces an early-fusion, multimodal approach to monitoring individual fishermen’s behaviour using wrist-worn inertial sensors. To characterise behaviour during fishing operations, accelerometer and gyroscope signals were collected onboard from 17 subjects across four behaviour classes. The proposed model performs channel-level early fusion to jointly process the heterogeneous six-channel time series, enhancing the extraction of complementary and discriminative representations from multi-sensor inputs. An experimental analysis was conducted, including a hyperparameter study, to assess the effects of class imbalance on model performance, compare the proposed approach with baseline and state-of-the-art models, and examine how different sensor configurations influence attention distribution within convolutional layers. Class imbalance affects model performance, with balanced training yielding 2.25 % higher accuracy than imbalanced training. A single accelerometer or gyroscope produces 8.43 % and 28.57 % lower accuracy than the fused accelerometer-gyroscope setup. The proposed model achieved the highest accuracy of 90.11 % with an inference time of 0.14 ms per sample compared with baseline and state-of-the-art models. Integration of wearable devices with human activity recognition algorithms is highly applicable to maritime scenarios and has strong potential to accurately distinguish fishermen’s behaviours, thereby providing a practical tool for monitoring human-related risk factors in maritime transportation.

Autores da comunidade :

• 

  C

  Carlos Guedes Soares

  ist11869

  Afiliações e fontes

  IST > CENTEC

  Centro de Engenharia e Tecnologia Naval e Oceânica

VoR - Versão publicada

Elsevier

https://www.sciencedirect.com/science/article/pii/S0924424725011574

Sensors and Actuators A: Physical

117351

398

0924-4247

other-engineering-and-technologies - Outras Ciências da Engenharia e Tecnologias

Palavras-chave

• Fishermen's behaviour recognition
• Wearable devices
• Fishermen's occupational safety

eng - Inglês