Abstract
In football penalty kicks, due to limited reaction speed and time requirements of the task, the goalkeeper must predict the direction of the ball’s flight before the opponent’s football touches. This article mainly focuses on the action recognition and tracking of penalty kick technology for football players on the football field. A system that integrates multiple sensors to solve the high requirements, privacy violations, and expensive equipment costs of scenarios and testing environments is proposed based on video image technology. The system collects 1,440 penalty action data from 24 subjects at the 11 m standard penalty spot using a 12 channel IMU embedded in the insole, and establishes a gait dataset. The trajectory analysis section extracts 30 segments from the competition video and manually marks 100 real motion trajectories for verification. The algorithm takes complementary filtering to fuse acceleration and angular velocity, combined with dual-mode convolutional neural network to achieve gait recognition, and completes trajectory tracking with Kalman filter as the core. The experimental results showed that the proposed algorithm achieved an area under the target tracking curve (AUC) of 91.15% in normal scenarios. In the case where members of the same team block the target, the maximum AUC was 67.73%, which was higher than that of comparative algorithms such as Multiple Instance Learning, Online AdaBoost, and Kernel Correlation Filter. This indicates that the proposed method is more accurate in target tracking than pure visual or single sensor schemes and has good robustness. In addition, the error between the estimated distance and the actual distance did not exceed 6%. The experimental results demonstrate that the proposed system breaks through the limitations of traditional video analysis in field and privacy, which can provide reliable data support for predicting penalty direction and optimizing actions in goalkeeper training.