Video event detection using auto-associative neural network and incremental SVM models

In this paper a new approach to video event detection is presented. This approach is based on HOG/HOF features optimized by an auto-associative neural network models for feature reduction and an incremental SVM model for event classification. This auto-associative neural network models are frequently used to reduce the size of feature vectors. In our approach, each event is modeled by a set of states, and each state is represented by a learning model containing a positive class (event) and a negative class (non-event). Experiments on real video sequences have shown encouraging results.