Action Pattern Analysis And Recognition Of Non-human Primates Under Controlled Environment

With the development of pattern recognition technology and the requirement for automatic analysis and recognition for animal related scenario, introducing the existing pattern recognition and machine learning technology into animal related research field has drawn the attention of researchers. However, due to the scarcity of the scenario and the limitation of data acquisition, the research of pattern analysis and recognition for animal related surveillance data has just began.Based on the above background, in this dissertation we use video surveillance of non-human primate (NHP) under controlled environment as research object. Our research goal is to automatically analyze and recognize NHP's behavior patterns. Through solving the emerging problems of symmetry pattern detection, behavior pattern clustering,pattern trajectory tracking, pattern measurement learning and pattern feature fusion in building the customized pattern recognition systems, we effectively mine the action pattern information useful for NHP related research from massive surveillance video so as to save manpower and improve work efficiency. This research is expected to play an active role in building animal surveillance related intelligent pattern analysis and recognition system in the future.The main contribution of this dissertation can be summarized as follows:1. To deal with the situation when video surveillance system monitors two objects simultaneously, we propose an adaptive symmetry detection algorithm based on local interest points. We transform the problem of region segmentation into symmetry pattern detection. The algorthm is improved by introducing a constraint on the amount of keypoints into local feature extraction,adding Ransac strategy to filter the matched pairs in feature pair matching, using geometric constraint in Hough space voting and utilizing the prior information symmetric pattern by Gaussian modeling. Experiments on NHP-Symmetry datasets show that the proposed algorithm is useful for segmenting the cage areas for NHP surveillance video. To further verify the generality of the proposed algorithm, we conduct experiments on public image datasets. Experimental results show that our algorithm is robust in extreme conditions like background noise, low contrast and smooth textures.2. We propose an action pattern clustering algorithm based on sticky hierarchical Dirichlet process based hidden Markov model (sticky HDP-HMM). This algorithm solves the problem of action pattern clustering in video sequences by taking it as segmentation and labeling of time series. Without defining the specific ation category, the algorithm labels the similar action patterns by the same labels in an unsupervised way. Therefore we can discover the distribution of NHPs' action patterns from the massive surveillance video, which is useful for further intellectual analysis of NHP surveillance video. Experimental results show the effectiveness of the proposed algorithm in mining the useful information in action patterns of NHP surveillance video.3. To deal with the specialty of monitoring objects and scenario in NHP video surveillance application, we propose a model-free pattern tracking algorithm for NHP's center of gravity. It can overcome the difficulties caused by occlusion and light changes.Experiments on NHP surveillance video dataset demonstrate the tacking accuracy and computational efficiency.4. We propose a pattern measurement learning algorithm based on adaptive local metric learning (L-LMNN). The proposed algorithm learns specific local metric according to the actual distribution of pattern data, which overcomes the diversity of data distribution and enhances the discrimination of pattern features in specific recognition task. Experiments on NHP surveillance video based action recognition demonstrate the effectiveness of the proposed local metric. To further verify the generality and extendibility of the proposed algorithm, we conduct experiments on Electroencephalograph (EEG) based personal identification.Experimental results show that our algorithm can improve the performance of EEG-based personal identification significantly.5. We propose a deep canonical correlation analysis based super vector (DCCA-SV) for action recognition. Compared with the existing feature encoding strategies, DCCA-SV embeds DCCA model into action feature distribution space. It encodes a pair of feature sets as one shared part and two separated parts so as to consider the correlations between different feature sets explicitly.Furthermore, the computation of DCCA-SV does not involve large scale matrix multiplication, which improves the efficiency in real application. Experiments on NHPAR and HMDB51 action recognition datasets show that the proposed algorithm has good performance in both accuracy and efficiency compared with the existing state-of-the-art methods. To further verify the generality of the proposed algorithm, DCCA-SV is extended to be the feature fusion step in EEG based personal identification. Experimental results show that our algorithm can improve the performance of EEG-based personal identification significantly, which verifies the generality of our feature fusion algorithm.