# Deep Learning Algorithms for Human Activity Recognition in Manual Material Handling Tasks

## TL;DR

This paper proposes and evaluates deep learning algorithms for recognizing manual material handling activities using wearable sensors, aiming to improve worker health and safety.

## Contribution

The study introduces and benchmarks four novel deep learning models for HAR in MMH tasks, emphasizing computational efficiency.

## Key findings

- BiLSTM and RCNN achieved high classification accuracy (95.7% and 95.9%) with lower computational complexity.
- DeepConvLSTM performed similarly but required significantly more operations, making it less efficient.
- BiLSTM and RCNN are suitable for embedded systems due to their lower computational demands.

## Abstract

Human Activity Recognition (HAR) is widely used for healthcare, but few works focus on Manual Material Handling (MMH) activities, despite their diffusion and impact on the workers’ health. We propose four Deep Learning algorithms for HAR in MMH: Bidirectional Long Short-Term Memory (BiLSTM), Sparse Denoising Autoencoder (Sp-DAE), Recurrent Sp-DAE, and Recurrent Convolutional Neural Network (RCNN). We explored different hyperparameter combinations to maximize the classification performance (F1-score,) using wearable sensors’ data gathered from 14 subjects. We investigated the best three-parameter combinations for each network using the full dataset to select the two best-performing networks, which were then compared using 14 datasets with increasing subject numerosity, 70–30% split, and Leave-One-Subject-Out (LOSO) validation, to evaluate whether they may perform better with a larger dataset. The benchmarking network DeepConvLSTM was tested on the full dataset. BiLSTM performs best in classification and complexity (95.7% 70–30% split; 90.3% LOSO). RCNN performed similarly (95.9%; 89.2%) with a positive trend with subject numerosity. DeepConvLSTM achieves similar classification performance (95.2%; 90.3%) but requires ×57.1 and ×31.3 more Multiply and ACcumulate (MAC) and ×100.8 and ×28.3 more Multiplication and Addition (MA) operations, which measure the complexity of the network’s inference process, than BiLSTM and RCNN, respectively. The BILSTM and RCNN perform close to DeepConvLSTM while being computationally lighter, fostering their use in embedded systems. Such lighter algorithms can be readily used in the automatic ergonomic and biomechanical risk assessment systems, enabling personalization of risk assessment and easing the adoption of safety measures in industrial practices involving MMH.