Skip to main content
. 2022 Jun 3;12:9282. doi: 10.1038/s41598-022-13220-2

Figure 1.

Figure 1

Overview of the proposed header detection algorithm. The video is divided into short segments and ball position is detected and tracked in each segment. Each frame is then cropped around the ball position. Spatiotemporal features of each cropped video segment are extracted using a convolutional neural network (CNN) and temporal shift module (TSM). The extracted feature vector is then classified as a header or nonheader event in the last step of the algorithm.