Behavior Detection In Complex Airport Surveillance Environment

In the field of computer vision and pattern recognition, behavior recognition in dynamic scenes is one of the most active research topics. It has a wide spectrum of promising applications. But most of the researches now are for the simple background, if the methods are applied in the complex scenes directly, the recognition rate is not ideal. So this paper proposes a method with the split model to get a better recognition rate.This paper is focused on behavior recognition in the airport. The amount of airport surveillance video and the video size are both impressive. The airport scene is very complex. It involves many moving objects, and they occlude each other badly. So this paper pays main attention to the behaviors close to the camera, and the others are ignored.This paper uses the temporal differencing for the object detection and the interest points as the behavior feature. Interest points refer to the points with a significant local variation of image intensities. The interest points reflect the local features of one behavior, and through the combination of these local features, one behavior can be distinguished from another. Harris corner method is employed for extracting the points. Based on the characteristics of airport video mentioned above, the research focus is how to gain the features of the needed behaviors in such a complex scene. It means that how to get the interest points located on one person. To solving this problem, K-means clustering arithmetic and Graphviz are used. With the help of human-computer interaction, most useful points are extracted, and we can get the split model.It is a key to describe a behavior effectively. A video sequence is represented as a bag-of-words using the features extracted above. This kind of representation requires creating a visual vocabulary. Then one behavior can be described as a histogram of visual word occurrences. Then SVM is applied to train the behavior model.During the testing stage, first of all, the features are extracted for the test sample. Then split model is applied to decide whether the interest points are needed to split. A split tree can be gained at last. Then cuboids and behavior descriptors are extracted for the leaves. At last, behavior model is applied for the behavior prediction. The recognition rate reaches 85.366%, and split model is the most important factor.