Revolutionizing Ergonomics in Agriculture Using AI

By Samudrapom DamReviewed by Susha Cheriyedath, M.Sc.Jan 10 2024

In an article recently published in the journal Applied Sciences, researchers proposed an artificial intelligence (AI)-powered posture classification system to effectively tackle ergonomic challenges agricultural workers face.

Background

Musculoskeletal disorders (MSDs) are one of the major occupational health concerns in the physically demanding agricultural sector. Several studies have indicated a high prevalence of MSDs and related symptoms among agricultural workers around the world.

In agriculture, ergonomic interventions can be critical in minimizing the MSD risk and improving overall worker safety and health. Factors that primarily contribute to repetitive movements and non-neutral postures can be identified and mitigated through ergonomic assessment.

Prolonged exposure to twisting and non-neutral postures, like awkward trunk bending, results in excessive stress on the MS system, which causes tissue damage and leads to the onset of MSDs. Thus, objective and accurate posture characterization and detection are critical to understanding ergonomic risks and implementing effective preventive measures.

The need for AI

In ergonomic studies, postural analysis has conventionally relied on subjective assessments, self-reporting, or manual observation, such as the Nordic Musculoskeletal Questionnaire and strain index. However, these approaches are prone to bias, time-consuming, and subjective.

The advent of AI, specifically computer vision techniques, has offered new opportunities to improve and automate posture characterization, including in the domain of ergonomics.These techniques use video and image processing algorithms to extract important features from the visual data captured using depth sensors or cameras. Different techniques, such as skeletal tracking, posture estimation, and image segmentation, have been utilized to analyze and identify body postures.

Researchers have evaluated deep learning (DL) and machine learning (ML) to identify and classify specific postures accurately based on sensor data. Specifically, transfer learning, a DL subfield, has gained significant attention as it can leverage pre-trained models to enhance classification performance.

Additionally, researchers have also leveraged other DL architectures and posture estimation models such as PoseNet and OpenPose to realize higher accuracy and effectiveness in detecting postures from image data. Among them, MoveNet is more suitable for application in different scenarios as it can identify the key joint features using information from the image. Studies demonstrated that MoveNet can be used to build software to monitor the physical activities of elderly individuals.

MoveNet can attain more precise results using only simple image data, specifically in scenarios where using complex and accurate measuring instruments is difficult for posture detection, which indicates the model’s significant application potential.

The MoveNet and classification-based approach

In this paper, the authors investigated the feasibility of using AI for posture detection in the agricultural field in the context of ergonomics. The objective of the study was to leverage ML and computer vision to overcome the limitations in robustness, accuracy, and real-time application that exist in conventional approaches such as direct measurement and observation.

Initially, field videos were collected in an outdoor plant nursery to capture the real-world scenarios of workers. Then, the trunk postures of workers were labeled into three different categories, including full forward bending, slight forward bending, and neutral.

Finally, the effectiveness of different posture detection approaches, including convolutional neural networks (CNNs), transfer learning, and MoveNet + classification, in accurately classifying trunk postures was investigated. Specifically, MoveNet was used to extract key anatomical features, which were subsequently fed into different classification algorithms, including artificial neural networks (ANN), random forests (RF), support vector machines (SVM), and decision trees (DT).

Overall, 200 images were used in this work. Additionally, the dataset was divided into a training dataset containing 85% of the data and a test dataset containing 15% of the data. The training dataset was formed using images obtained from the external data source, while the test dataset was formed using images obtained from the agriculture field during this research and images from the external data source. Accuracy, precision, recall/sensitivity, and F1-score were used as evaluation metrics to comprehensively assess the model’s overall performance and its ability to accurately classify various postures.

Research findings

Although a 92.5% training accuracy was observed when the CNN model was applied to trunk posture classification in the ergonomics context, the testing accuracy was only 41.2%, which indicated that the model could not effectively generalize to unseen data and suffered from substantial overfitting. No significant improvement in results was observed even when the overfitting was reduced through data augmentation.

The transfer learning models displayed better generalization capabilities and attained higher testing accuracies compared to the standard CNN model, with MobileNet achieving the highest testing accuracy of 65.56% among all evaluated transfer learning models. However, the testing accuracies of these models were still low.

Researchers obtained the best performance using MoveNet together with ANN. They achieved 87.41% F1-score, 87.52% recall, 87.46% precision, and 87.80% testing accuracy using MoveNet + ANN. MoveNet + RF showed the second-best performance with 85.36% testing accuracy.

To summarize, the findings of this study demonstrated the feasibility of using AI-driven precise posture classification systems to improve safety prevention practices and worker health in the agricultural industry.