The asymmetry angle is the angle between the sagittal line and the asymmetry line (see Fig. 1), measured at both the start and end of a lift [1]. It is assumed that at the start, the body aligns with the asymmetry line, and the feet remain stationary before the torso rotates. Similarly, at the end, the body aligns with the asymmetry line, and the feet remain stationary before releasing the object. The mid-sagittal line is defined as the line passing through the midpoint between the inner ankle bones (considered the body center) and lying in the mid-sagittal plane, based on the neutral body position in the applications manual for the RNLE [1] for the ease of operational measurements. Trunk rotation (lift asymmetry) occurs after lifting the object with both hands. The asymmetry line, defining the asymmetry plane, is determined by projecting the object’s center on the floor (point of projection) to the body center.

Calculation of the asymmetry angle (A) requires determining the body center (O), the point of projection (P), and a point (Q) as shown in Fig. 1. According to [6], the x–y coordinates of the body center O were estimated as the mid-point between the two ankle joints, and the point of projection P was estimated using the x–y coordinates of the mid-point of the knuckles of both hands. A second point on the mid-sagittal line (point Q) is estimated as the mid-point of the second distal phalanges (second toes) of both feet of the subject. Given the x-y coordinates of O, P and Q, A may be computed using the following formula: (2)AMoCap=cos−1OP→⋅OQ→‖OP→‖⋅‖OQ→‖ where OP→ and OQ→ are two dimensional vectors in the x–y plane from body center O to P and Q, respectively. ‖OP→‖ and ‖OQ→‖ are the magnitudes of the corresponding vectors.

We redefine the sagittal plane as the vertical plane intersecting the midpoint between hip landmarks instead of the traditional ankle joints. From an ergonomics perspective, the rationale is that the sagittal plane is perpendicular to the hips when the upright body is in the neutral anatomical position and these landmarks are not affected by rotation of the feet. For CV applications, hips are preferable to ankles due to their more reliable estimations in the field. This is because ankles, as extremity joints, present fewer constraints and are therefore harder to estimate [38], and are more prone to occlusion, which degrades accuracy [39]. Additionally, we opt for wrist joints over knuckles, as they are anatomically close and commonly detected by CV algorithms [21].

A skeletal joint body model of a worker performing a lift is shown on the left-hand side of Fig. 2. Denote the projection of the 3-D coordinates of the hip landmarks on the floor as H_L and H_R, the projection of the 3-D coordinates of the wrist landmarks on the floor as W_L and W_R as shown in the right side of the Fig. 2. The body center here is denoted by m_h, the point of projection is now denoted by m_c. A unit-vector vs→ that is perpendicular to the line segment HRHL¯ and originates from m_h gives the direction of the mid-sagittal line. Similarly, a unit-vector va→ from m_h to m_cgives the direction of the asymmetry line. Then, the CV angle of asymmetry (A_CV ) can be estimated as the arc cosine of the inner product of these two unit-vectors (3)ACV=cos−1(v→aTv→s).

Given a single view video, the 3-D coordinates will need to be estimated from the video using another approach, which will be described in Section II-D.

The same landmarks used for A_CV were also calculated using the corresponding 3-D MoCap measurements. We term this quantity Ā_CV. Specifically, Ā_CV was calculated using the hip landmarks on the floor as H_L and H_R, the projection of the 3-D coordinates of the wrist landmarks on the floor as W_L and W_R measured from the MoCap system.

CV algorithms enable automatic estimation of a 3-D human skeleton from monocular video. Initially, body location is detected in each video frame using a human detector, resulting in a bounding box of the body region [24]. Next, a 2-D pose estimator, HR-Net, processes each cropped human image to estimate 2-D coordinates for 17 skeletal joints [25]. Tracking then links 2-D landmarks belonging to the same individual across frames [26]. Within a local window around the lifting frame, VideoPose3D [20], a 3-D pose estimator, processes tracked 2-D landmarks to estimate the relative depth of each 3-D landmark for the lifting video frame, as depicted in Fig. 3.

The VideoPose3D algorithm [20] needs to be fine-tuned with domain-specific data to enhance its performance [14], [15], [27]. This dataset, described in detail in the subsequent section, comprises video clips of subjects performing lifting tasks and corresponding 3-D joint coordinates measured using MoCap. Initially, HR-Net is employed on each frame to estimate the 2-D poses (joint locations). Then, the pre-trained VideoPose3D model is refined using HR-net estimated 2-D poses as the input and the corresponding MoCap-derived 3-D joint locations as targets.

Since MoCap-derived 3-D landmarks are in a different coordinate system from that of the video camera, alignment is necessary using a PnP algorithm [23]. The PnP algorithm estimates the camera pose, namely the extrinsic parameters of the video camera, using the set of 2-D and corresponding 3-D joint coordinates. Then, the MoCap-measured 3-D joint positions are transformed into the camera coordinate system for refinement training. To validate the alignment process, MoCap-derived 3-D joint coordinates are back-projected onto the video camera’s image plane using the estimated camera pose. The normalized reprojection error, indicating the difference between back projected 2-D joint coordinates and those estimated using HRNet, divided by the diagonal length of video frames, is 0.02. Given a video frame resolution of 640 by 480 pixels, the average back projection error is approximately 16 pixels. During refinement of the VideoPose3D model, we employed leave-one-subject-out cross-validation (see Section II-G.) to mitigate concerns of overfitting. Hyperparameters selected for fine-tuning include a learning rate of 0.001, batch size of 64, training epochs of 20, and the AMSGrad optimizer [28].

The video data utilized in this article originated from a previous NIOSH study conducted in Cincinnati, Ohio [8]. Ten NIOSH associates participated in this prior study, performing 12 lifting tasks under IRB approval, with informed consent obtained from each participant. Hand locations for each lifting task were adjusted based on the 12 lifting zones defined by the American conference of governmental industrial hygienists (ACGIH) threshold limit value (TLV) for lifting to encompass a wide range of lifting motions [29]. Each subject repeated each lifting task three times, resulting in a total of 36 randomly assigned lifting trials completed in one session. During each trial, subjects walked from a starting point to one of 12 marked locations, lifted a wire grid frame weighing 0.45 kg, turned around, walked to shelf height to drop off the frame, and returned to the finish line. The lightweight frame ensured clear visibility of MoCap markers.

Among the 360 video recordings, 147 exhibited trunk rotation before stepping or significant hip movement, while the remaining videos had near-zero asymmetry angles A. These videos were used for analysis as a control condition for zero asymmetry angles. Asymmetry angle A was measured for three types of lifts (symmetric, asymmetric toward the camera, and asymmetric away from the camera) after lift initiation (BOL), as depicted in Fig. 4.

Ethical approval was obtained from the appropriate local ethics committee or Institutional Review Board (University of Wisconsin-Madison Federal Wide Assurance no. #FWA00005399), protocol number 2016-0297-CR006, and informed consent has been obtained.

The machine-learning algorithm underwent leave-one-subject-out cross-validation to evaluate its accuracy and robustness while mitigating potential overfitting [27]. The dataset was divided into ten subsets, and this process was repeated ten times. In each iteration, one subject was designated as the test set, and the remaining nine subjects constituted the training set. This led to the creation of ten models, each trained and evaluated using a different test set. The overall performance, measured by angle estimation error in degrees, was averaged across these ten models.

Evaluation metrics included the mean absolute error and standard deviation of the estimated asymmetry angle compared to ground truth MoCap values [8], [9]. To complement the mean absolute error, a threshold accuracy with a 15° error interval was introduced (see Section III-A), following the RNLE Applications Manual [22]. Precision was assessed by comparing the 2-D CV-measured asymmetry angle (A_CV) against the asymmetry angle computed using the same MoCap-measured landmarks (Ā_CV). Accuracy was evaluated by comparing the 2-D CV-measured asymmetry angle (A_CV) against the conventional 3-D MoCap-derived angle (A_MoCap).

The CV-measured asymmetry angle (A_CV) was compared against the asymmetry angle, computed using the same landmarks measured by MoCap (Ā_CV). We term this hip-to-hip error as both used hip joints. The overall mean absolute angular error was 6.25° (SD = 10.19°). Across all subjects, the minimum mean error was 4.58° (SD = 5.86°) and the maximum mean error was 8.59° (SD = 11.28°).

From (1), a 15° increment of A results in a 5% reduction of AM and hence a 5% change in RWL. We consider less than a 5% change in the RWL within the practical limits of precision of any of the variables included in the RNLE. Consequently, we evaluated the percentage of samples whose absolute error was less than 15°. The results showed that 91% of samples had an error of less than 15° across ten subjects.

A histogram across all subjects, for the number of estimated hip-to-hip error cases plotted against error range, is shown in Fig. 5. The histogram confirmed that most of the errors were small and within the range of [−15°, 15°] (91%), yet had a few outliers having errors as large as −60°. We summarize the percentiles of precision error above specific thresholds in Table I using a 15° interval following the RNLE guidelines. The percentiles for errors beyond 30° and 45° were 3.1% and 1.1%, respectively.

As shown in Fig. 4, there were three types of lifts: symmetric lift (SL); turn toward the camera (TT); and turn away from the camera (TA). The mean absolute angular error and percent within 15° precision and percent within 30° precision for each type of lift are summarized in Table II. In state-of-the-art 2-D pose estimation algorithms, the 2-D coordinates of joints that are not observed directly are often estimated using prior information, leading to lower accuracy of the 2-D body joint positions. This leads to inaccurate 3-D body joint position estimates and hence higher angle estimation error. This explains why in Table II, the “turn away” lift scenario exhibits higher error rate compared to the SL and TT scenarios.

The CV measured asymmetry angle A_CV was further compared to the conventional 3-D MoCap derived angle (A_MoCap). We term this quantity hip-to-ankle error as A_CV used hip while A_MoCap used ankles. The overall mean absolute accuracy error was 9.45° (SD = 14.01°). Across all subjects, the minimum mean error was 5.31° (SD = 7.34°) and maximum mean error was 16.11° (SD = 16.62°).

A histogram across all subjects, for the number of estimated accuracy error cases plotted against error range, is shown in Fig. 6. The histogram confirmed that most of the errors were small and within the range of [−15°, 15°] (79%), yet had a few outliers having errors as large as −80°.

We summarize the percentiles of accuracy error above specific thresholds in Table III using a 15° interval. The percentiles for errors beyond 30° and 45° were 6.7% and 1.9% respectively. The mean absolute accuracy error and percent within 15° accuracy and percent within 30° accuracy for each type of lift are given in Table IV.

The primary distinction between A_CV and A_MoCap lies in the method used to define the mid-sagittal line. In the CV method, hip orientation determines this line, while the MoCap method relies on the toes of the feet. Given that the hips initiate movement before the feet during the lifting trials’ turning phase, the difference between A_CV and A_CV would likely be smaller than between A_CV and A_MoCap. If the hips and toes rotate at the same rate, A_CV and A_MoCap angles would become the same. The vector perpendicular to the hip, which CV employs to calculate the sagittal line, closely follows the point of projection, offering a more realistic representation than the method used in A_MoCap.

In [30], manual measurements of the asymmetry angle showed a mean error of −4.5° (SD = 4.8°), compared to our method’s −2.34° (SD = 10.19°). Despite the slightly lower SD in manual measurements, the error difference is within 15°, likely having minimal impact on RWL calculations. The key benefit of our approach over manual measurements is not accuracy, but its non-intrusive nature, enabling potential for fully autonomous monitoring and broadening its use beyond standard workplace tasks.