Chinese Journal of Computers   Full Text
Title	Trajectory Tracking and Recognition Using Bi-Directional Nonlinear Learning
Authors	HU Zhao-Hua1),2) FAN Xin2) LIANG De-Qun2) SONG Yao-Liang1)
Address	1)(School of Electronic Engineering and Optoelectronic Technology, Nanjing University of Science and Technology, Nanjing 210094) 2)(School of Information Engineering, Dalian Maritime University, Dalian, Liaoning 116026)
Year	2007
Issue	No.8(1389—1397)
Abstract &   Background	Abstract Object trajectory is one of the most important cues for tracking and behavior recognition and can be widely applied to numerous such as visual surveillance and guidance. However, it is a difficult problem to directly model spatio-temporal variations of trajectories due to their high dimensionality and nonlinearity. This paper proposes a novel trajectory tracking and recognition algorithm by combining a bi-directional deep neural network called "autoencoder" into a particle filter. First, the "autoencoder" network embeds the high-dimensional trajectories in a two-dimensional plane based on a peculiar training rule and learns a trajectory generative model by the inverse mapping. Then a series of plausible trajectories are generated by the trajectory generative model. In the tracking process, the generated samples from the plausible trajectory set are weighted by the color likelihood and are resampled so as to obtain target state estimation at each time step. Finally the tracking trajectory is recognized by min-distance classification method in the two-dimensional plane. In particular, the "autoencoder" provides such a bi-directional mapping between the high-dimensional trajectory space and the low-dimensional space and is therefore able to overcome the inherited deficiency of most nonlinear dimensionality reduction methods (e.g. LLE and ISOMAP) that do not have an inverse mapping. The experiments on tracking and recognizing handwritten digits show that the proposed algorithm can robustly track and exactly recognize in background clutter. keywords autoencoder network; trajectory generative model; nonlinear dimensionality reduction; object tracking Background This work is supported by the Sub-Project of the National Key Technologies R&D Program of China on Image-based Intelligent Traffic Management under grant No.2004BA111B01. Object tracking and recognition is one of the key issues on intelligent video surveillance and management. Object trajectory provides an important cue for tracking and behavior recognition. This work addresses the fundamental issues to exploit the trajectory cue into tracking and recognition, i.e., how to represent the variations and uncertainties of trajectories and how to incorporate this model into the framework of tracking and recognition. The authors have developed the techniques to localize and recognize license plates from static images under the support of the grant. It is believed that the work provide a good starting point to immigrate the previous algorithms on static images to dynamic video sequence applications. However, it is a difficult problem to directly model spatio-temporal variations of trajectories due to their high dimensionality and nonlinearity. To deal with large amounts of high-dimensional data, many researchers try to discover the latent structure of high-dimensional data. However, most existing methods, e.g., LLE and ISOMAP, merely find out a latent geometric structure of the data which preserves certain relationship between the data points of an available training set. They does not provide an explicit bi-directional mapping between the high-dimensional data space to the low-dimensional embedded space. To realize the mapping, these nonlinear dimensionality reduction methods have to resort to additional complicated techniques such as RBF (radial basis function) and GPLVM (Gaussian Process Latent Variable Models). Considering above deficiencies, the authors introduce the "autoencoder" network which can convert high-dimensional data to low-dimensional codes by training a neural network with with multiple hidden layers. The "autoencoder" provides such a bi-directional mapping between the high-dimensional trajectory space and low-dimensional latent space and is therefore able to overcome the inherited deficiency of most nonlinear dimensionality reduction methods. Moreover, the trained network implies a generative model for plausible trajectories which can be readily combined into a Bayesian framework for tracking and recognition. In this work, the authors develop the target tracking and recognition system by integrating the "autoencoder" and the particle filter. First, the "autoencoder" network embeds the high-dimensional trajectories in a two-dimensional plane based on a peculiar training rule and learns a trajectory generative model by the inverse mapping. Then a series of plausible trajectories are generated by the trajectory generative model. In the tracking process, the generated samples from the plausible trajectory set are weighted by the color likelihood and are resampled so as to obtain target state estimation at each time step. Finally the tracking trajectory is recognized by min-distance classification method in the two-dimensional plane. Background Intrusion detection system (IDS) deals with huge amount of data which contains irrelevant and redundant features causing slow training and testing process, higher resource consumption as well as poor detection rate. Existing studies to build lightweight IDS have proposed two main approaches: parameters optimization of classification algorithms and feature selection of audit data. Feature selection is one of the key topics in IDS. Feature selection involves finding a subset of features to improve prediction accuracy or decrease the size of the structure without significantly decreasing prediction accuracy of the classifier built using only the selected features. Methods for feature selection have been essentially divided into two categories: filter methods and wrapper methods. Wrapper methods generally perform better than filter methods, but they involve some more computational complexity and require more execution time than the filter methods. Some researchers have proposed hybrid feature selection methods which combine wrapper and filter methods. However, the number of selected features is large and the performances of intrusion detection system which based on hybrid feature selection are not perfect. Current research results show that wrapper feature selection algorithm performs better than other two methods, except its computational complexity. Therefore, this paper proposes a novel wrapper-based feature selection algorithm to build lightweight IDS. The approach is able not only to solve the computational complexity but also to guarantee high detection rates. The researches of this paper are supported in part by the National Basic Research Program(973 Program) of China under grant No.2004CB318109 and the National Information Security Project of China under grant No.2005C39. The former project is to develop a security analysis of network systems. The latter is to build lightweight IDS. The work in this paper is part of the research on building lightweight IDS, trying to use novel feature selection algorithm toward building lightweight IDS.