Research on application of athlete gesture tracking algorithms based on deep learning
- PDF / 1,390,064 Bytes
- 9 Pages / 595.276 x 790.866 pts Page_size
- 88 Downloads / 143 Views
ORIGINAL RESEARCH
Research on application of athlete gesture tracking algorithms based on deep learning Teng Long1 Received: 18 August 2019 / Accepted: 6 November 2019 © Springer-Verlag GmbH Germany, part of Springer Nature 2019
Abstract It is difficult to track the posture of players in the course, mainly because of the changing environment and players. In this paper, the improved neural network is used to extract the trajectory characteristics of the athletes in the football player’s game video, and the network is trained on a large number of data objects containing similarity objects, which improves the ability of the algorithm to distinguish the athlete’s trajectory. A scheme of soccer attitude tracking based on twin neural network. The experimental results show that the algorithm has a good effect in the field of football and the accuracy is over 90% and the Siamese neural network is better than a traditional convolutional neural network. Keywords Football players tracking · Improved neural network · Deep learning
1 Introduction In modern life, football is very popular among public audience. Player tracking is of very great practical values and realistic significance. In the past decade, the tracking algorithm has made great progress (Gao et al. 2014, pp 188–203; Wang and Lu 2014; pp 3478–3485, Zhang et al. 2015), it is still a big challenge to design a robust tracker which can deal with serious shielding and significant appearance changes. In recent years, target tracking has been widely researched as an important topic in the field of computer vision. Among those objects being tracked, tracking of human postures has become a hot topic in tracking research fields. Human posture tracking technology plays a vital role in applications such as video monitoring (Wang and Yeung 2013). In a given video and initial frame region, human pose tracking technology is to continuously find the region of human pose in the subsequent frames of the initial frame (Cui et al. 2016). For object tracking, a lot of researchers have proposed effective methods. Henriques (2015, pp 583–596) designed a Kernel Correlation Filter (KCF) and used it to train a discriminant classifier, the training and classification were * Teng Long [email protected] 1
School of Sports and Arts, Hunan University of Chinese Medicine, Changsha, Hunan, China
carried out through sample generation with cyclic matrixes. DLT and SO-DLT used offline training of a depth model with auxiliary image data and conducted fine adjustment during online tracking. FCNT and DeepSRDCF extracted characteristics with a CNN network which had been trained in advance on a large scale of classification datasets, so as to solve the problem of lack in training data. MDNet made pre-training of CNN with videos instead of images, so as to obtain universal expression ability of targets. RTT (Henriques 2015, pp 583–596) established a multi-directional recurrent neural network model to explore reliable target parts useful for tracking. At present, human body posture tracking can be
Data Loading...
