The detection and localization of the ball in sport videos is crucial to better understand events and actions occurring in those sports. Despite recent advances in the field of object detection, the automatic detection of balls remains a challenging task due to the unsteady nature of balls in images. In this paper, we address the detection of small, fast-moving balls in sport video data and introduce a real-time ball detection approach based on the YOLOv3 object detection model. We apply specific adjustments to the network architecture and training process in order to enhance the detection accuracy and speed: We facilitate an efficient integration of motion information, avoiding a complex modification of the network architecture. Furthermore, we present a customized detection approach that is designed to primarily focus on the detection of small objects. We integrate domain-specific knowledge to adapt image pre-processing and a data augmentation strategy that takes advantage of the special features of balls in images in order to improve the generalization ability of the detection network. We demonstrate that the general trade-off between detection speed and accuracy of the YOLOv3 model can be enhanced in consideration of domain-specific prior knowledge.

Motion visualization is an attractive way to provide support for a range of recreational and competitive sports. In skateboarding, sensor technology in particular can help visualization systems capture the motion of athletes to provide relevant information to athletes, judges, and spectators. This article describes the authors' proposed application of a 9D inertial-magnetic measurement unit (IMMU) based real-time trick classification and visualization system. It also reports on a survey they conducted with skateboarders that asked about the usefulness, acceptance, and future ideas of such a system.

The Kalman filter (KF) is an extremely powerful and versatile tool for signal processing that has been applied extensively in various fields. We introduce a novel Kalman-based analysis procedure that encompasses robustness towards outliers, Kalman smoothing and real-time conversion from non-uniformly sampled inputs to a constant output rate. These features have been mostly treated independently, so that not all of their benefits could be exploited at the same time. Here, we present a coherent analysis procedure that combines the aforementioned features and their benefits. To facilitate utilization of the proposed methodology and to ensure optimal performance, we also introduce a procedure to calculate all necessary parameters. Thereby, we substantially expand the versatility of one of the most widely-used filtering approaches, taking full advantage of its most prevalent extensions. The applicability and superior performance of the proposed methods are demonstrated using simulated and real data. The possible areas of applications for the presented analysis procedure range from movement analysis over medical imaging, brain-computer interfaces to robot navigation or meteorological studies.

With tracking systems becoming more widespread in sports research and regular training and competitions, more data are available for sports analytics and performance prediction. We analyzed 2523 ski jumps from 205 athletes on five venues. For every jump, the dataset includes the 3D trajectory, 3D velocity, skis’ orientation, and metadata such as wind, starting gate, and ski jumping hill data. Using this dataset, we aimed to predict the expected jump length (xLength) inspired by the expected goals metric in soccer (xG). We evaluate the performance of a fully connected neural network, a convolutional neural network (CNN), a long short-term memory (LSTM), and a ResNet architecture to estimate the xLength. For the prediction of the jump length one second after take-off, we achieve a mean absolute error (MAE) of 5.3 $\mathrm{m}$ for the generalization to new athletes and an MAE of 5.9 $\mathrm{m}$ for the generalization to new ski jumping hills using ResNet architectures. Additionally, we investigated the influence of the input time after the take-off on the predictions’ accuracy. As expected, the MAE becomes smaller with longer inputs. Due to the real-time transmission of the sensor’s data, xLength can be updated during the flight phase and used in live TV broadcasting. xLength could also be used as an analysis tool for experts to quantify the quality of the take-off and flight phases.