Study And Application Of Animal Behavior Automatic Analysis Based On Posture Recognition

Assessment of animal behavioral effects is widely used in researches of central nervous functions, such as motorial function, learning and memory. It has become an essential part of neural science, especially neural physiology and ethopharmacology etc. Methods and technologies of behavioral analysis have progressed from manual behavioral observation and transducer-detection to video tracking technology. Recently, digital image processing technique has been applied to tracking the position and path of an animal, and also further locomotory parameters such as distance moved, average speed. Though this technique has more advantages than former methods, it is based only on analysis of locomotory patterns with the animal represented as spatial coordinates of its barycenter. However, behavior consists of postures and changes of postures (including changes of positions). As an essential part of behavior, posture contains more psychological information. Consequently, recognition of animal postures and detailed description of animal behavior have become an interesting research area in ethology, especially in behavior researches of rodents.By attacking several key steps on behavior analysis, we developed some algorithms and a new automatic behavior analysis system for real-time posture recognition, ethogram description, and online analysis.In order to be robust in the environment of animal behavior experiment, such as illumination changes, a new object detection method based on RGB threshold subtraction was presented. In contrast to the image processing methods used in current automatic analysis systems, it showed much higher immunity to environmental influence.Rat postures were categorized into five classes for empirical reason, i.e., groom, stretched attend, stationary rotation, sit, and rear. During experiment, rats could move freely in one area, which can bring its orientation and position changes constantly. Relying on this, feature of the shape of rat image boundaries should berotation, translation invariant. Normalized curvature was used to describe the contour of rat. A novel posture recognition method based on contour curvature and hierarchical clustering analysis was then presented.Curvature is a key notion in intuitionistic description of contour. A new curvature estimation method based on chain-code was introduced and applied to calculate the curvature function of a rat's contour. And then the spectra of the curvature functions were used as eigenvectors for posture classification. According to multiformity of rats' postures, subclasses in each posture class were constructed using hierarchical clustering analysis in order to obtain higher correct recognition rates. In our experiment, two groups of images with the same resolution (DPI: Dot per inch: 19.4) were used as training set and validation set, and the correct recognition rates are 94.16% and 89.58% respectively.In real application, resolution changes of images may be induced by rat size variety and camera focus changes and so on. When images with different resolution ( DPI: 11.6) was used as validation set, the correct recognition rate is 63.5%, which showed that the presented method is not applicable for resolution changes. In order to be robust in resolution changes, four "cursory" features of rat contour were extracted, which were also proved to be rotation, translation, and scaling (RTS) invariant. The features were "cursory", which result in high immunity to intra-class multiformity. A naive Bayes classifier is a probabilistic classifier. Although it has a much simpler structure than other models, it is still observed to be successful as a classifier in practice. Another shape analysis method based on Naive Bayes classifier was presented and was selected to generalize images of different target object resolution (DPI: Dot Per Inch), with the "cursory" features as attribute values. In our experiment, five groups of image frames with different resolution were acquired. One group (DPI: 19.4) was used as training set to estimate the possibility density of each...