Catálogo de publicaciones - libros

Localization of facial feature points is an important step for many subsequent facial image analysis tasks. In this paper, we proposed a new coarse-to-fine method for extracting 20 facial feature points from image sequences. In particular, the Viola-Jones face detection method is extended to detect small-scale facial components with wide shape variations, and linear Kalman filters are used to smoothly track the feature points by handling detection errors and head rotations. The proposed method achieved higher than 90% detection rate when tested on the BioID face database and the FG-NET facial expression database. Moreover, our method shows robust performance against the variation of face resolutions and facial expressions.

This paper presents a new Bayesian face tracking method under particle filter framework. First, two adaptive feature models are proposed to extract face features from image sequences. Then the robustness of face tracking is reinforced via building a local dual closed loop model (LDCLM). Meanwhile, trajectory analysis, which helps to avoid unnecessary restarting of detection module, is introduced to keep tracked faces’ identity as consistent as possible. Experimental results demonstrate the efficacy of our method.

Though recognizing human action from video is important to applications like visual surveillance, some hurdles still slower the progress of action recognition. One of the main difficulties is view dependency, and this causes the degeneration of many recognition algorithms. In this paper, we propose a template-based view-independent human action recognition approach. The action template comprises a series of “action hyperspheres” in a nonlinear subspace and encodes multi-view information of several typical human actions to facilitate the view-independent recognition. Given an input action from video, we first compute the Motion History Image (MHI) and corresponding polar feature according to the extracted human silhouettes; recognition is achieved by evaluating the distances between the embedding of the polar feature and the virtual centers of the hyperspheres. Experiments show that our approach maintains high recognition accuracy in free viewpoints, and is more computationally efficient compared with classical NN approach.

Adaptive background subtraction (ABS) is a fundamental step for foreground object detection in many real-time video surveillance systems. In many ABS methods, a pixel-based statistical model is used for the background and each pixel is updated online to adapt to various background changes. As a result, heavy computation and memory consumption are required. In this paper, we propose an efficient methodology for implementation of ABS algorithms based on multi-resolution background modelling and sequential sampling for updating background. Experiments and quantitative evaluation are conducted on two open data sets (PETS2001 and PETS2006) and scenarios captured in some public places, and some results are included. Our results have shown that the proposed method requires a significant reduction in memory and CPU usage, meanwhile maintaining a similar foreground segmentation performance as compared with the corresponding single resolution methods.

This paper presents a novel frame-layer rate control technique that adaptively determines the frame complexity for bit allocation in order to satisfy the target bit-rate constraints without degrading the decoded video significantly. To do this, we first obtain the edge energy of each frame to measure the frame complexity as well as to determine the weighting of a frame for bit allocation. We then present a new bit-rate traffic model for bit allocation to achieve a better conformance to the target bit-rate. Finally, we integrate the edge energy complexity measure into the rate-quantization (R-Q) model. Our results shows robust improvements over the current rate control methods adopted in H.264/AVC in terms of meeting the target bit-rate as well as determining the quality of the decoded video.

The brand-new SP-frame in H.264 facilitates drift-free bitstream switching. Notwithstanding the guarantee of seamless switching, the cost is the bulky size of secondary SP-frames. This induces a significant amount of additional space or bandwidth for storage or transmission. For this reason, a new motion estimation and compensation technique, which is operated in the quantized transform (QDCT) domain, is designed for coding secondary SP-frames in this paper. So far, much investigation has been conducted to evaluate the trade off between the relative sizes of primary and secondary SP-frames by adjusting the quantization parameters. But, our proposed work aims at keeping the secondary SP-frames as small as possible without affecting the size of primary SP-frames by incorporating QDCT-domain motion estimation and compensation in the secondary SP-frame coding. Simulation results demonstrate that the size of secondary SP-frames can be reduced remarkably.

Intra mode selection and motion estimation for spatial and temporal prediction play important roles for achieving high video compression ratio in the latest video coding standards, such as H.264/AVC. However, both components take most of the computational cost in the video encoding process. In this paper, we propose an efficient intra mode prediction algorithm based on using the mode conditional probability learned from a large amount of training video sequences with the ground truth modes of each block to be encoded and its neighboring block modes as well as its associated image content features. By applying the proposed intra-mode selection algorithm into the H.264 reference code, we show significant reduction of the computation time with negligible video quality degradation for H.264 video encoding.

In this paper, we propose fast mode decision algorithms for both intra prediction and inter prediction in H.264. In order to select the candidate modes for intra4x4 and intra16x16 prediction efficiently, we have used the spatial correlation and directional information. We have also applied an early block size selection method to reduce the searching time further. The fast inter mode decision is achieved by an early SKIP mode decision method, and a fast mode decision method for 16x16 and P8x8 modes. Extensive simulations on different test sequences demonstrate a considerable speed up by saving the encoding time up to 82% for intra prediction and 77% for inter prediction on average, compared to the H.264 standard, respectively. This is achieved at the cost of negligible loss in PSNR values and small increase in bit rates.

H.264 achieves considerable higher coding efficiency compared with previous video coding standards, whereas the complexity is increased significantly. This paper proposed a novel fast algorithm based on structural similarity (SSIM) in motion estimation (ME) process (FMEBSS), which can greatly reduced the complexity of ME by eliminating the unnecessary search positions and reducing complex prediction modes. Simulation results demonstrate that the proposed method can averagely reduce the coding time by about 50%, and compression ratio is improved at the same time, while the degradation in video quality is negligible.

In this paper, a novel shot boundary detection algorithm is introduced for H.264/AVC bitstreams. This algorithm relies on features present in a compressed video bitstream, i.e., macroblock types and reference directions. In contrast to existing algorithms, the proposed technique is able to analyze bitstreams with frames containing different types of slices. To deal with such frames, formulas are presented that work on inter- and intra-coded slices. The results obtained for the different types of slices are combined by calculating a linear combination, taking into account their size. This way, a metric is found that can be used for the automatic detection of shot boundaries. Results show that the proposed algorithm has a high accuracy in terms of recall and precision for video sequences with frames containing multiple slice types.