Reidentifying soccer players in broadcast videos using Body Feature Alignment Based on Pose

Abstract

Re-identification (re-id) of people in images is a well-studied problem in computer vision for many applications. The re-identification of players in broadcast videos of team sports is the main subject of this work. We specifically concentrate on recognizing the same player in images taken at any given time during a match from various camera angles. Some significant differences exist between this task and other traditional person re-id applications, such as same team wear highly similar clothes, for each identification, there are only a small number of samples, and low resolutions of the images. One of the most difficult problems in object re-identification is extracting robust feature representation (ReID). Even though methods based on convolution neural networks (CNNs) have had significant success. But to improve extracting features, we present the novel approach Body Feature Alignment Based on Pose, utilizing pose landmarks to extract the image's useful information. During the feature constructing stage, our method makes use of human landmarks to obtain the angles and distances between the joints. According the results, the proposed method provide comparable improvements for convolutional networks.