Ding Xiaoqing

Full body tracking-based human action recognition

In this paper, we present a novel method for human action recognition with the combined global movement feature and local configuration feature. The human action is represented as a sequence of joints in the 4D spatio-temporal space, and modeled by two HMMs, a conventional HMM for global movement feature and an exemplar-based HMM for configuration feature. Firstly, an adaptive particle filter is adopted to track the marker-less actor¿s 3D joints. Then, the combined features are extracted from the full body tracking results. Finally, the actions are classified by fusing two HMMs. The effectiveness of the proposed algorithm is demonstrated with experiments on 7 actions by 12 actors. The results prove robustness of the proposed method with respect to viewpoints and actors.

Background subtraction based on a combination of texture, color and intensity

A new background subtraction algorithm based on a combination of texture, color and intensity information is presented. The texture is depicted by DLBP, a modified version of LBP, color is a local template in HS color space and intensity the pixel intensity value. The approach works robustly both on rich texture areas and uniform areas even with noise and can deal with weak cast shadows effectively. The comparative experiments show that the overall performance of our method is better than other methods in most scenes.

Multiple candidate characters in the post-processing for off-line handwritten Chinese character recognition

The size of candidate character set is very important in the contextual post-processing for Chinese character recognition, which should be changed with the scripts recognition accuracy of isolated Chinese character recognition. In this paper, based on the confidence theory of candidate character, two methods are proposed to automatically decide the size of candidate character set. One is called static decision method in light of the scripts overall situation, and the other is called dynamic decision method in allusion to each candidate character set in the script. Experimental results on off-line handwritten Chinese character recognition show the effectiveness of the proposed methods.

A new kind of neuron model with a tunable activation function and its applications

A new neuron model with a tunable activation function, denoted by the TAF model, and its application are addressed. The activation function as well as the connection weights of the neuron model can be adjusted in the training process. The two-spiral problem was used as an example to show how to deduce the adjustable activation function required, and how to construct and train the network by the use of thea priori knowledge of the problem. Due to the incorporation of constraints knowna priori into the activation function, many novel aspects are revealed, such as small network size, fast learning and good performances. It is believed that the introduction of the new neuron model will pave a new way in ANN studies.A new neuron model with a tunable activation function, denoted by the TAF model, and its application are addressed. The activation function as well as the connection weights of the neuron model can be adjusted in the training process. The two-spiral problem was used as an example to show how to deduce the adjustable activation function required, and how to construct and train the network by the use of thea priori knowledge of the problem. Due to the incorporation of constraints knowna priori into the activation function, many novel aspects are revealed, such as small network size, fast learning and good performances. It is believed that the introduction of the new neuron model will pave a new way in ANN studies.

Unified HMM-based layout analysis framework and algorithm

To manipulate the layout analysis problem for complex or irregular document image, a Unified HMM-based Layout Analysis Framework is presented in this paper. Based on the multi-resolution wavelet analysis results of the document image, we use HMM method in both inner-scale image model and trans-scale context model to classify the pixel region properties, such as text, picture or background. In each scale, a HMM direct segmentation method is used to get better inner-scale classification result. Then another HMM method is used to fuse the inner-scale result in each scale and then get better final segmentation result. The optimized algorithm uses a stop rule in the coarse to fine multi-scale segmentation process, so the speed is improved remarkably. Experiments prove the efficiency of proposed algorithm.

Fast Text Matching in Digital Videos

This paper proposes a fast and reliable video text matching algorithm suitable to tracking and locating text in given videos. The core of the algorithm relies on using fast cross-correlation and pyramid down-sampling techniques in a coarse-to-fine scheme. Fast local correlation, referred to as single matching phase, is achieved by using recursive computation schemes, which enabled us to minimize the amount of calculations required at every new pixel. By working with rectangular sub-images, which obtained by segmenting the frames at different levels of the pyramid, the speed of the algorithm can be increased and the intermediate memory storage requirement is reduced. We have tested our match algorithm in a large set of experiments with video clips recorded from television and achieved good matching results.

Fast Text Registration and Enhancement in Digital Video

This paper addresses a fast and reliable video text registration and enhancement algorithm suitable to tracking text and optical character recognition (OCR). The fast registration algorithm based on the normalized cross-correlation and pyramid structure in a coarse-to-fine scheme. After extracting reference text block, we use the text registration method to find the corresponding text blocks in next consecutive frames. By using intensity-hue-saturation (IHS) fusion method to enhance text block. This proposed algorithm working with rectangular sub-images, which obtained by down-sampling the frames at different levels of the pyramid, the speed of the algorithm can be increased. We have tested our match algorithm in a large set of experiments with video clips recorded from television and achieved good matching results.

Adaptive particle filter with body part segmentation for full body tracking

This paper presents a novel approach for marker-less 3D full body pose tracking using adaptive particle filter. Firstly, the search space decomposition strategy and body part segmentation method are used to reduce the calculation complexity due to the large degrees of freedom. Then an adaptive particle filter is adopted to track each body part. This new technique is a significant improvement over the standard particle filter with the advantage of adaptive particle number for each body part. Experimental results on tracking several challenging action sequences have shown that the proposed 3D full body tracker is able to effectively handle rapid no-linear movements, large changes of viewpoint, and different actors. The average errors of joint position are from 0.56 to 1.13 voxel in these action sequences.

Recognition of 3-D objects based on Markov random field models

The recognition of 3-D objects is quite a difficult task for computer vision systems. This paper presents a new object framework, which utilizes densely sampled grids with different resolutions to represent the local information of the input image. A Markov random field model is then created to model the geometric distribution of the object key nodes. Flexible matching, which aims to find the accurate correspondence map between the key points of two images, is performed by combining the local similarities and the geometric relations together using the highest confidence first method. Afterwards, a global similarity is calculated for object recognition. Experimental results on Coil-100 object database, which consists of 7200 images of 100 objects, are presented. When the numbers of templates vary from 4, 8, 18 to 36 for each object, and the remaining images compose the test sets, the object recognition rates are 95.75%, 99.30%, 100.0% and 100.0%, respectively. The excellent recognition performance is much better than those of the other cited references, which indicates that our approach is well-suited for appearance-based object recognition.

Open-Set Face Verification Algorithm Using Competitive Negative Samples

A novel face verification algorithm using competitive negative samples is proposed. In the algorithm, the tested face matches not only with the claimed client face but also with competitive negative samples, and all the matching scores are combined to make a final decision. Based on the algorithm, three schemes, including closest-negative-sample scheme, all-negative-sample scheme, and closest-few-negative-sample scheme, are designed. They are tested and compared with the traditional similarity-based verification approach on several databases with different features and classifiers. Experiments demonstrate that the three schemes reduce the verification error rate by 25.15%, 30.24%, and 30.97%, on average, respectively.