New technology has entered the world of prevention to assist and identify movement patterns and tasks that can be modified to help decrease musculoskeletal injuries and protect workers. Cameras have become one of the most prominent tools used in movement prevention. Learning about the features and possible weaknesses of such technology helps educate safety professionals and employers to decide what systems are best suited to help them meet their goals.
Simple observation and opinion-based assessment of the worker performing the task can be sufficient to identify the obvious movements that may increase a worker’s injury risk, such as bending and twisting their back to lift an object rather than keeping their back straight and bending their knees. These assessments focus primarily on the quantity of the movement.
However, it’s often the unknown injury risks that cannot be assessed through observation and opinion that can significantly reduce the worker’s injury risk. These injury risks are caused by the quality of movement, such as how smooth and controlled a movement is compared to faster, jerky movements. It is much harder to assess the quality of the movement by observation and opinion, which is why the use of movement analysis tools in the workplace is increasing across the Occupational Health and Safety industry.
One of the most popular tools involves the use of a single video camera and analysis of the footage using computer vision. However, the level of accuracy for this method needs to be evaluated to determine how effective it can be for safety professionals.
Benefits of using single–camera video analysis and computer vision to measure worker movements for injury risk assessment
The most accurate method of human movement analysis requires a motion capture system, which can measure the three-dimensional position and movement of the body parts in all three planes (sagittal, frontal, and transverse). This requires a minimum of nine electromagnetic or reflective markers positioned over the body and a minimum of nine specialized cameras to detect the position of these markers.6 Though this method is highly accurate, it is not practical as it requires a lot of equipment and cables, which take a long time to set up, is limited to the small area where the equipment is set up, takes a long time to position the markers on the person. It also requires a certain level of expertise to use the specialized software and analyze the data. As a result, motion capture systems are rarely used outside of a lab-based environment and are very time-consuming and expensive.
Therefore, other methods have evolved that are more cost-effective, have user-friendly equipment and software, and are capable of being used in a variety of environments (including worksites). The two main methods that are growing in popularity include the combined use of video and wearable sensors and single-camera computer vision.
The combination of wearable sensors and video maintains a high level of accuracy, as the sensors are still capable of measuring the movements of the body and body parts in all three planes of movement, and the video can provide a visual representation of these movements.6 The main sources of error with this method include the position and location of the sensors on the body (as this can impact the reliability of data between reports) and the quality of the sensors (as this can reduce the accuracy of the measurements). So, it’s important that the combined wearable sensor and video solution has been validated by an independent body, such as a university or professional association.
The least accurate movement analysis method involves the use of single-camera video analysis through Computer Vision (CV) technology. Since this method requires nothing more than a smartphone app, it is gaining popularity due to its ease of use. However, users of this technology need to be aware of the high risk of error using single-camera video analysis.
How does Computer Vision work?
The position and orientation (pose) of the body and body segments must be established to measure the joint angles and limb movements of a worker through a camera video. This is done through modeling, which is a software function that uses certain criteria to estimate the position of the body and body segments in relation to each other and the environment for each frame of the video. These criteria include body configuration (e.g. legs attached to the pelvis, pelvis attached to the spine etc.), body shape and appearance, and camera viewpoint.
For accurate analysis, these criteria should be known in advance. For example, establishing a fixed camera viewpoint (angle to the plane of movement being assessed and distance from the body and body parts) and known body part dimensions (body width, limb lengths, etc.). Most CV platforms do not collect this information as part of the analysis process, so the model needs to provide an estimate. These estimates produce many false positives, as there are often many images that look very similar to body parts. For example, suppose a safety professional was assessing whether a work task involved hazardous arm movements and there was a curved surface on a piece of equipment behind the shoulder of the person performing the task.
The CV model may identify the curved surface as the shoulder. This would then lead to inaccurate measurements of the shoulder position and movement, which may lead to a hazardous position or movement not being identified. Therefore, due to the variations between different people in shape and appearance, the lack of consistency with camera position (angle and distance from the body), and the environment, the CV model is the first step in single-camera movement analysis where error can occur.
Parallax Error: The biggest source of error in single–camera analysis.
Also referred to as perspective error or scaling error, parallax error occurs when the plane of movement being assessed is not perpendicular to the camera. This can cause errors when measuring joint angles and body position (quantity of movement), as well as inaccurate measurement of motion (quality of movement). For example, if you used CV to measure the angle of an arm that is reaching in a direction towards the camera and then measured the same reaching movement but with the camera positioned perpendicular to the direction the arm is reaching, the measurements would be very different. The images below demonstrate the comparison between the CV analysis of the same body position from three different angles. This demonstrates how the reaching shoulder (right shoulder) and lower back angle measurement of the same pose from single–camera CV can be very different depending on the camera angle.
However, the most significant parallax error occurs when measuring the quality of movement through acceleration and velocity of movements. Objects moving at a constant speed across the frame will always appear to move faster if they are closer to the camera or slower if they are further away from the camera. Therefore, an accurate assessment of the stress on the worker’s body due to fast, jerky movements is not possible unless more than one camera is used.
Numerous studies have been conducted to investigate the parallax errors associated with single-camera movement analysis. The only valid and reliable studies that demonstrated the data from single camera movement analysis involved lower limb gait analysis on a treadmill, positioned perpendicular to the limb movements being assessed (in the sagittal plane) and the background was standardized.1,2,3,4 When gait measurements were made in any other plane of movement (frontal or transverse), the level of accuracy decreased significantly.5
The research that focused on the analysis of whole body movements demonstrated that the data was only accurate and reliable when the video was perpendicular to the plane of movement and markers were placed on anatomical landmarks to reduce the error from the CV modelling.7,8
Consequences of error in single–camera movement analysis
If you’re responsible for assessing the injury risks associated with specific work tasks, the biggest risk of error is measuring a false negative. This will occur when the data incorrectly identifies a low injury risk when the injury risk is actually high. Motion parallax error is the main cause of this type of error, as fast, jerky movements that should be avoided may be missed by the single-camera video analysis.
False positives are the other concern when using single-camera video analysis, as you may end up assessing work tasks as high injury risk when they actually have a moderate or low injury risk. The best way to avoid these errors and ensure your assessments are as accurate as possible is to combine video movement analysis with measurements from wearable sensors.
New technologies offer innovative ways to improve movement safety for employers, safety teams, and employees. It is advantageous to assess different movement capture options when considering what would work best.
Learn more about injury prevention technology, and how to accurately capture human movement to lower risk.
References
- Mark F Reinking et. al. Reliability Of Two-Dimensional Video-Based Running Gait Analysis. Int J Sports Phys Ther. 2018 Jun;13(3): 453-461.
- Bart Dingenen et.al. Test-retest reliability of two-dimensional video analysis during running. Phys Ther Sport. 2018 Sep:33: 40-47.
- Camma Damsted et.al. Reliability Of Video‐Based Quantification Of The Knee‐ And Hip Angle At Foot Strike During Running. Int J Sports Phys Ther. 2015 Apr; 10(2): 147–154.
- Alexandria Michelini et.al Two-dimensional video gait analysis: A systematic review of reliability, validity, and best practice considerations. Prosthetics and Orthotics International. Volume 44, Issue 4.
- Thiago Jambo Alves Lopes et.al. Reliability and Validity of Frontal Plane Kinematics of the Trunk and Lower Extremity Measured With 2-Dimensional Cameras During Athletic Tasks: A Systematic Review With Meta-analysis. J Orthop Sports Phys Ther. 2018 Oct;48(10):812-822.
- Benjamin R. Hindle et.al. Inertial-Based Human Motion Capture: A Technical Summary of Current Processing Methodologies for Spatiotemporal and Kinematic Measures. Applied Bionics and Biomechanics. 2021; Published online 2021 Mar 26.
- Chris Ugbolue et.al. The evaluation of an inexpensive, 2D, video based gait assessment system for clinical use. Gait & Posture. Volume 38, Issue 3, July 2013, Pages 483-489
- Bart Dingenen et.al. The reliability and validity of the measurement of lateral trunk motion in two-dimensional video analysis during unipodal functional screening tests in elite female athletes. Phys Ther Sport. 2014 May;15(2):117-23.
