Learning to Recognize Complex Moving Objects
复杂运动目标的学习与识别

This paper presents an automatic system for human face and moving object recognition. The system developed is based on a novel recurrent stochastic neural network, it has a strong learning power and is able to recognize a moving target in real time. The detection of the moving object is implemented by utilizing the skin color distribution and the motion information. The object is tracked in real time with an efficient adaptive mean shift algorithm. The work in this paper is mainly focused on the disign of the novel recurrent neural network and the efficient incremental Boltzmann learning algorithm. The improved simulated annealing technique is also discussed. Theoretical results offer a unique solution to the training of a large size network. Experiments on human face recognition are carried out with a recurrent neural network of 4827 neurons and 129951 connections. The results show the performance of the recognizer is comparable to that of the well known TrueFace system.