Counterattack Detection in Broadcast Soccer Videos using Camera Motion Estimation Mohamad-Hoseyn Sigari, Hamid Soltanian-Zadeh Control and Intelligent Processing Center of Excellence School of Electrical and Computer Engineering College of Engineering University of Tehran Tehran, Iran hoseyn_sigari@yahoo.com, hszadeh@ut.ac.ir Vahid Kiani, Hamid-Reza Pourreza Machine Vision Research Laboratory Computer Engineering Department Faculty of Engineering Ferdowsi University of Mashhad Mashhad, Iran vahid.kiani@rocketmail.com, hpourreza@um.ac.ir Abstract—This paper presents a new method for counterattack detection using estimated camera motion and evaluates some classification methods to detect this event. To this end, video is partitioned to shots and view type of each shot is recognized first. Then, relative pan of the camera during far-view and medium- view shots is estimated. After weighting of pan value of each frame according to the type of shots, the video is partitioned to motion segments. Then, motion segments are refined to achieve better results. Finally, the features extracted from consecutive motion segments are investigated for counterattack detection. We propose two methods for counterattack detection: (1) rule-based (heuristic rules) and (2) SVM-based. Experiments show that the SVM classifier with linear or RBF kernel results in the best results. Index Terms—Broadcast soccer video, counterattack detection, camera motion estimation, event detection, video analysis. I. INTRODUCTION Nowadays, a variety of digital videos such as movies, news and sport videos are available on the web, hard disks, and non- volatile memories. Thus, traditional indexing/retrieval methods (tag-based search), manual analysis and summarization of video data are boring and challenging. Thus, we need some automatic or semi-automatic systems to analyze the video and extract content and semantic of video. Sport video, especially soccer video is one of the interesting video types for analysis and information extraction. Sport video analysis is performed for video classification [1], video summarization [2, 3, 4, 5], video retrieval [6, 7], team/player tactic analysis [8, 9], developing referee-assistant systems [10, 11] and etc. Thus, it is demanded by coaches and sport reporters for technical and tactical analysis [8, 9] or statistical information of a match (e.g. percentage of ball possession [12]). Additionally, TV networks are interested to show advertisements on broadcast sport videos using virtual advertising techniques [13]. Almost all of the mentioned applications of sport video analysis are based on event detection. In fact, event is the basic entity of a video in the semantic level of computational hierarchy. Thus, a robust and high performance analysis is usually performed based on detection of events occurred in the video. Soccer event detection systems can be categorized into two categories. The first category includes the systems detect events in broadcast videos. These systems usually used for the application of video summarization [2, 3, 5], video indexing/retrieval [6, 7] and virtual advertising [13]. On the other hand, the second category includes the systems developed for event detection using a network of special cameras. The cameras used by such systems are fixed and usually covers the entire field of match. For example, the systems developed for real-time detection of goal vent [10] and offside event [11] are categorized in the second category. Such systems are usually used as referee-assistant systems. However researchers proposed a lot of methods for event detection in broadcast soccer videos, but they usually consider the most important events including goal [2, 3, 14, 15, 16], attack [2, 7, 14, 16] and corner [2, 14, 16, 17]. There are only a few researches that detect other events such as offside [2, 14], penalty [16], card [2] and counterattack [16] events. This paper presents a new method for counterattack detection. In sports science, there are two types of soccer counterattacks. One is slow and based on ball controlling that relies on a lot of short soccer passes in all directions. The other one is a quick direct attack by moving the ball forward into the goal of the other team. In general, counterattack refers to the second definition. Thus, in this paper we assume that a counterattack is a quick attack into the opposition half by the defending team after winning the ball from the team previously attacking. Counterattack detection may be used for different applications such as tactical analysis, video indexing/retrieval and video summarization. There is a main cue to detect a counter attack: fast turnover from one side of the playing field to the other side. This cue is also used for counterattack detection in [16]. The method proposed in [16], uses finite state machine and some heuristic rules to model some basic events (e.g. forward pass and turnover) in the feature space. Then, the system detects a counterattack when two specific basic events, i.e. forward pass and turnover are detected consequently. According to their  ,QWHUQDWLRQDO 6\PSRVLXP RQ $UWLILFLDO ,QWHOOLJHQFH DQG 6LJQDO 3URFHVVLQJ $,63   ,((( 