Fall Detection for Shipboard Seafarers Based on Optimized BlazePose and LSTM

Area of Science:

• Maritime Safety
• Computer Vision
• Artificial Intelligence

Background:

• Seafarer falls pose significant risks, necessitating timely medical assistance.
• Existing fall detection systems lack accuracy and real-time performance for maritime environments.

Purpose of the Study:

• To develop an accurate and real-time fall detection algorithm for seafarers.
• To enhance maritime safety by enabling prompt warnings following a fall.

Main Methods:

• Utilized the BlazePose network for human keypoint extraction from video.
• Implemented a novel head detector based on Vitruvian theory and an offset vector for bounding box acquisition.
• Employed a Long Short-Term Memory (LSTM) neural network for fall behavior detection.
• Enriched datasets with URFall and FDD public datasets for training and validation.

Main Results:

• Achieved 100% accuracy and 98.5% specificity in detecting seafarer falls.
• Demonstrated real-time detection capabilities with a frame rate of 29 fps on a CPU.
• The algorithm shows strong generalization ability and practical applicability.

Conclusions:

• The BlazePose-LSTM algorithm offers a practical and effective solution for seafarer fall detection.
• The proposed method can be deployed on standard vision sensors, enhancing onboard safety systems.
• This technology has the potential to significantly reduce seafarer injuries and fatalities due to falls.