Highlight Goal Event Detection In Soccer Videos

The soccer videos are characterized as a long duration and huge data. Owing to thefacts that the network transmission bandwidth is limited and that the audiences arealways interested in highlight events, automatic detection of highlight events in soccervideos has become a research hotspot in the field of sports video analysis. At present,the common mainstream methods mainly include two categories: the machine-learningmethods and the methods based on manual rules. The former suffers complex semanticevent models and strict requirements of sufficient and typical sample data in modeltraining; the latter faces a considerable cost of manual efforts in establishing thesemantic manual rules and a poor performance of event detection. Therefore, how toconstruct semantic event models with good performance, establish simple and effectivesemantic rules, and realize semantic event detection accurately and completely is theresearch difficulty in the current sports video field. Focusing on the problem ofhighlight goal event detection in soccer videos, this paper proposes two methods.(1) A novel method based on HMM (Hidden Markov Model) and the semantic rulefor goal event detection. By semantic shot annotation, the video clip is described as asequence composed of long, medium, close-up, spectator, and replay shots. Byconsidering soccer domain knowledge, the HMM model for the goal event isconstructed and the goal event detection based on HMM is achieved. Based on thecontent analysis of the goal and non-goal clips, a new shot feature, namely semanticobservation weight, is defined. The normalized semantic weighted sum rule isestablished by using the new feature and goal event detection based on semantic rules isrealized. Finally the two detection results are fused by optimal weights in the decisionlevel by using the weighted fusion scheme based on logic distance, and the goal eventdetection is completed. Experimental results show that the presented method achieves96.43%precision and100%recall for goal event detection in soccer videos, whichoutperforms traditional HMM and semantic rule methods.(2) A new framework based on HCRF (Hidden Conditional Random Field) andaffective semantics for goal event detection. The HCRF model for real-time shotannotation is built which can realize the annotation of multiple types of semantic shotssimultaneously. The semantic shot sequence of the video is obtained. Based on the shot feature of semantic observation weight, the affective arousal model is constructed tointerpret the inherent affective semantics from the perspective of affective semantics.The affective arousal value sequence of the video is obtained. Under the condition ofsmall training data set, a simple HCRF model for goal event detection is effectivelybuilt based on the mapping relationship between the goal event and the sequences ofvideo semantic shots and affective arousal values. The inherent pattern of goal events isexcavated from the perspectives of both video structure semantics and affectivesemantics. The goal event detection is precisely achieved. Experimental results showthat this method achieves a good performance in goal event detection.Finally, the summary of this paper is given and the future direction of the researchis presented.