Behavior classification is based on behavior representation, and further determines the category of behavior according to the extracted behavior characteristics. Behavior classification methods can be roughly divided into direct classification methods and state space classification methods. The direct classification method does not consider the time relationship in the image sequence, and directly converts a single frame image or image sequence into a feature vector to classify behavioral patterns. The state-space law defines each static posture as a state, and the switching between various states satisfies a certain probability relationship. The behavior sequence is regarded as a traversal process between these different states, and various traversal processes are calculated. The corresponding joint probability, and its maximum value is used as the basis for the classification of behavior patterns.

 

Direct classification methods mainly include k-nearest neighbor (k-NN) classifier, SVM, RVM, Boosting, etc. State space classification methods mainly include hidden Markov model (hidden Markov model, HMM), dynamic Bayesian network (dynamic Bayesian network, DBN), conditional random field (conditional random field, RcF) and dynamic time warping (dynamic time warping, DTW) etc. They are usually insensitive to small changes in time or space of behavior sequences, and are suitable for modeling and analyzing complex behaviors.

 

There is also a class of behavior recognition called anomalous behavior detection. Different from the classification and recognition of various human behaviors, the purpose of abnormal behavior detection is to detect behaviors different from normal behaviors from sequence images. Due to the diversity and scene dependence of behaviors, abnormal behaviors are usually defined as behaviors that seldom or never appear in normal videos, and the specific categories of abnormal behaviors are often not the focus of attention. The process of abnormal behavior detection generally includes feature extraction, model training, and anomaly detection.