Automatic Detection of Faults in Race Walking: A Comparative Analysis of Machine-Learning Algorithms Fed with Inertial Sensor Data

Abstract

The validity of results in race walking is often questioned due to subjective decisions in the detection of faults. This study aims to compare machine-learning algorithms fed with data gathered from inertial sensors placed on lower-limb segments to define the best-performing classifiers for the automatic detection of illegal steps. Eight race walkers were enrolled and linear accelerations and angular velocities related to pelvis, thighs, shanks, and feet were acquired by seven inertial sensors. The experimental protocol consisted of two repetitions of three laps of 250 m, one performed with regular race walking, one with loss-of-contact faults, and one with knee-bent faults. The performance of 108 classifiers was evaluated in terms of accuracy, recall, precision, F1-score, and goodness index. Generally, linear accelerations revealed themselves as more characteristic with respect to the angular velocities. Among classifiers, those based on the support vector machine (SVM) were the most accurate. In particular, the quadratic SVM fed with shank linear accelerations was the best-performing classifier, with an F1-score and a goodness index equal to 0.89 and 0.11, respectively. The results open the possibility of using a wearable device for automatic detection of faults in race walking competition.

Keywords: activity recognition; illegal steps; inertial sensors; machine-learning algorithms; race walking.

Figure 1

Example of a support vector machine (SVM). Circles and squares represent features related to two different classes. The distance 2/‖w‖ is the defined margin.

Figure 2

Example of a k-nearest neighbor (kNN) with the number of neighbors set to 1 or 10. k represents the computed distance. Circles and squares represent features related to two different classes. Star represents new data to classify.

Figure 3

Scheme of an artificial neural network (ANN) classification. n, m, and c represent the number of input, hidden, and output layers, respectively. w and b represent the weights and biases related to the interconnections among neurons.

Figure 4

Scheme of a generic decision tree (DT). t_i represent the thresholds of the splitting. Grey rectangle represents the input, red ones represent the outputs, and blue ones represent the decision nodes where the rules of the classification are applied.

Figure 5

Placement of IMUs (orange probes) on an athlete during the experimental procedure.

Figure 6

Flow chart for the identification of the best-performing classifiers. RW stands for race walking condition.

Figure 7

Confusion matrix related to the ${}_{\mathrm{a}}{}^{\mathrm{SH}}\mathrm{SVM}_{\mathrm{q}}$ considering all subjects together. 1, 2, and 3 stand for regular, loss of contact, and knee-bent condition, respectively. Numbers in green cells represent the correct classifications, while those in red cells are misclassifications.

Figure 8

Confusion matrix related to the simplified ${}_{\mathrm{a}}{}^{\mathrm{SH}}\mathrm{SVM}_{\mathrm{q}}$ considering all subjects together. 1 and 2 stand for regular and irregular race walking, respectively. Numbers in green cells represent the correct classifications, while those in red cells are misclassifications.