Research On Event Detection For Basketball Video Using Features Of Superimposed Objects And Its Application To Physical Education

This paper addresses the problem of event detection for basketball video by using features extracted from superimposed objects. Light-weight and effective event detection algorithms can be developed for basketball video based on direct relations between the superimposed features and certain event semantics. However reading (localizing and recognizing) these superimposed objects is challenging due to complex background, low resolution, small size and blur. To overcome the challenge, this paper proposes two novel event detection algorithms for basketball video. By exploiting more discriminative temporal features rather than static image features for reading superimposed objects, the proposed algorithms are more accurate than traditional algorithms on event detection. The research of this paper includes following five aspects.The first aspect of research is localization and recognition for single video clock (The video contains only one digital clock and it keeps running properly). Based on the ideas previously proposed by our group, this research makes a complete work and proposes an end-to-end algorithm for reading single video clock. Compared with traditional algorithms, the proposed algorithm localizes clock digits based on the periodicity of pixel color change in second digit region, which can replace the tedious and error prone image processing procedure. In this research we also build a dataset for testing our clock reading algorithm. The dataset is composed of one thousand videos and publicly available.The second aspect of research is localization and recognition of two clocks for basketball video. In our previous work an algorithm for reading single video clock (there is only one digital clock in video and the clock keeps running properly) has been proposed. However reading clocks for basketball video poses different difficulties to us since the two clocks are independent, plus they pause and reset frequently. To overcome challenges caused by these difficulties, this paper proposes an end-to-end clocks reading algorithm for basketball video, which is composed of algorithm for single video clock reading and two new methods. The first is a new set of functions for instantiating the pixel periodicity method for detecting pixels in second regions of clocks. This new method does not need to detect the clock transit frames. The second is a procedure for finding the proper digit-sequence from the second place of the clock. This technique solves the difficulty caused by the frequent pauses and the resets of two clocks. The extracted digit sequence then can be recognized via sequence template matching method proposed in the previous work.The third aspect of research is suspension boundary detection for basketball video. With the information extracted from the game clock and the shot clock this paper infers the start and the end points of suspension clips. Experimental results show that the proposed algorithm can get much more accurate boundary points of suspensions than the traditional algorithms can do at a much lower computing cost.The fourth aspect of research is scoring type recognition for basketball video. This paper divides scoring type detection to two sub-problems, which are free throw scoring detection and three-points scoring detection. For free throw scoring detection, this paper proposes a new superimposed feature to judge whether a scoring event is free throw. And there is no need to recognize the score digit when extracting this feature. For three-points scoring detection, this paper proposes a CRF(Conditional Random Field) model for score digit recognition and judge if a scoring event gains three points by recognizing score digits before and after the event. Experimental results show that the proposed algorithm can achieve higher accuracy than traditional algorithms can do.The fifth aspect of research is the development of a function module for multimedia basketball teaching assistant system. The function is scoring event searching and replay for basketball video. Based on the proposed event detection algorithms for basketball video in this paper, the developed function module can accurately detect the scoring events and corresponding segments in basketball video. These detected video segments can be used as presenting resources in basketball teaching and learning activities.