Quantitative analysis of swimming technique

On this page you will find a joint project of the University of Kassel - Institute of Sport and Sport Science and the University of Augsburg - Chair of Machine Vision and Multimedia. The aim of this project is to develop empirically derived technique models for the sport of swimming and to use these for performance diagnosis and prognosis. Special software for markerless recognition of swimmers' joint positions is used for this purpose (Lienhart et al., 2014) and is intended to simplify the complex and difficult movement analysis in swimming. At the Olympic training centers, video material on performance tests in the swimming channel is available from the successful years in German swimming, which covers a period of several years. This includes successful movement techniques that have been proven to be associated with top performances in swimming. It seems sensible to summarize the most important common movement characteristics of the top swimmers of the past in technique models in order to be able to provide improved feedback for the current generation of swimmers. In the service research project carried out here, the segment movements of successful swimmers from Germany are quantitatively analyzed by swimming position using automatic recognition of the joint points. The data is then analyzed for common characteristics using a factor analysis method. Based on the movement patterns obtained from the factor analysis, a mathematical-statistical description of a technique model is created.

This project is funded by the Federal Institute for Sports Science (BISp project ZMVI4-072041/18-19).

In this section you will find useful information on the topic: Compliance with the technical guidelines.

As soon as relevant information is available for all types of swimming, the content will be added.

For the differentiation between men and women in the backstroke position, the gender of the swimmer can be derived from the first component of the PCA via the top swimmers with an accuracy of 73.3%. Two men and two women from a group of 15 videos were misclassified. This is remarkable insofar as there are no gender-differentiated swimming technique models in the specialist literature.

If the technical possibilities of linear discriminant analysis are now transferred to a differentiation in distance lengths (and the associated swimming speeds) between 100m and 200m, the distance length can be derived from the technical characteristics of the movement with a 92.3% probability, and this across genders. This means that, for the first time, movement techniques can now be determined as a function of swimming speed and gender on the basis of top German swimmers. This makes it possible to visualize distance and gender-specific technical characteristics for the swimmers and present them in a video format.

In this section, the development of digital adaptive technique models is now shown, using the butterfly and backstroke as examples. For butterfly swimming, 13 videos of four men and three women and for backstroke swimming 14 videos of four men and four women were used. All swimmers had participated in at least one adult world championship and were recorded at the speed they were practicing at the time. Two cycles with a constant swimming speed were selected from the swimmers' individual poses, which were combined to form a cyclical swimming movement. Due to different recording times, the videos were available in 25 Hz (*.avi codec) and 29.97 Hz (*.mp4 codec), so that a time normalization was first carried out for the manually post-processed coordinate data. A principal component analysis (PCA) was then calculated for each individual swimmer. PCA offers the possibility of compressing information and reducing it to essential characteristics. The analysis of the segment coordinates of a cyclic swimming movement described below is based on the procedure used by Troje (2002) to determine male and female gait patterns.

From the fully automatic tracking procedure of the Augsburg Chair of Multimedia and Machine Vision and the manual post-correction, body segment coordinates are available for each swimmer for each pose of two selected cycles. These contain the position information for 14 relevant body segment endpoints (ankle, knee, hip, wrist, elbow, shoulders, in the middle of the neck and the cranial end of the head). The position of each segment endpoint is determined by the x and y coordinates for each video frame, so that each individual pose can be represented as a 28-dimensional vector. Since the butterfly swim and the backstroke take about 2 to 3 seconds over two cycles, the recording frequencies of the videos result in 100 to 180 individual images and thus 100 to 180 individual poses for the respective swimmer.

With the help of an initial PCA, the swimming movements in the respective position were reduced to the most striking movement characteristics for each swimmer across their individual poses and relevant technical (main) components were extracted. For this purpose, the eigenvectors and eigenvalues of a matrix consisting of the individual poses were determined over two cycles (M = 28 x number of individual poses). This revealed the five principal components shown in Figure 2 for the two swimming positions, which together explain over 98% of the total variance of the pose sequences.

In our analysis, five components were included to determine the individual swimming technique. The five main components we included can explain 98.93% of the information content of the individual swimming technique on average for the butterfly swimming position and 98.36% for the backstroke swimming position. This means that the swimming technique can be reproduced across genders with a probability of almost 99%. The associated error is smaller than the variation in swimming technique found in 100m freestyle races for men and women (Seifert et al., 2007).

Since swimming is a cyclical movement skill, the temporal behavior of the five components can be modeled using simple sine functions (Troje, 2002). Each sine function is characterized by its frequency or period (b), its amplitude (a) and its phase (c). On this basis, function values can be derived for all five principal components, indicating the temporal behavior of the mean swimming motion with a high agreement (97.4%). Our analysis was based on a double sinusoidal function with six parameters:

a1*(sin(b1*x+c1)+a2*(sin(b2*x+c2)

The collection of these function values represents a further step in data reduction, as the eigenvalues of the principal components can be replaced by only six parameters per principal component when considering the previous function.

In a second step, a matrix structure/table is formed in which the charge coefficients and the function values (which can approximate the principal components using the listed function) are stored for each float. Based on this matrix, a vector is created that contains information on 1) the position of the swimmer, 2) the individual movement technique and 3) the associated temporal patterns. The special feature of this representation is that it can be transformed. In this way, an "average swimming technique" can be derived, which is based on the individual movement patterns and characterizes the greatest possible commonality between top swimmers.

In the third step of the technique analysis, the matrix structure is used to identify further information from the movement techniques. A further principal component analysis is used for this purpose. Only the normalized values from the previous matrix structure/table of swimmers are included in a further factor analysis. The most important criteria for differences in movement technique can now be identified on the basis of the average movement technique.

Overall, the procedures just described and the associated derivation of technique characteristics can be used to identify cross-class differences between different characteristics in swimming technique. By means of a subsequent linear discriminant analysis, it is possible to test the differences between the technique characteristics derived above (Table 2), whereby relevant and irrelevant characteristics can be highlighted. Discriminant analysis is used as a multivariate method, which is ideal for classification. To differentiate between groups, further information (e.g. year the video was recorded, gender, swimming speed, height, body weight) must be added to the matrix structure. Subgroups can then be classified for each of these incoming variables using linear discriminant analysis, whereby the technical characteristics for these subgroups are output in the form of parameters.

• Approach
• Documenta
• Conference Program

FAQ:

Armin Kibele
akibele(at)uni-kassel.de

University of Kassel
Institute for Sport and Sport Science
Damaschkestr. 25
34121 Kassel

Phone: +49(0)561-804-5397

Matthias Weigelt
matthias.weigelt(at)uni-paderborn.de

Paderborn University
Faculty of Natural Sciences
Warburger Str. 100
33098 Paderborn

Phone: +49(0)5251-60-3200