Search

Action recognition from videos is important due to its numerous applications in real life scenarios. Most commonly action recognition is used for surveillance but it is also used in other applications such as entertainment and gaming. Moreover, due to high demand of automation and easy availability of higher processing power, researchers are aiming for better ways to accurately classify human actions using videos. Different methods have been proposed in literature in order to identify actions. However, most of these methods suffer from limitations such as implementation complexity and inability to recognize complex actions. In this paper, we propose a novel action recognition technique which is easy to implement and is robust to complex actions. While there are different methods to capture motion, trajectories based methods have proven to be an excellent way to represent motion in videos. Inspired by that, we sampled dense points in different frames of the video to track down their displacement information to encode the trajectories. We used dense trajectories along with oriented gradients and motion boundary information as features. We use codebook generated using Gaussian mixture model and Fisher vector to encode our features. Finally, after reducing the dimensionality of our data, we use linear Support Vector Machine to identify the actions. The experimental evaluations performed on a recent benchmark action recognition dataset showed that our method is able to achieve encouraging results compared to other competing methods.

This paper presents a novel cross-view gait recognition technique based on the spatiotemporal characteristics of human motion. We propose a deep fully-connected neural network with unsupervised learning which transfers the gait descriptors from multiple views to the single canonical view. The proposed non-linear network learns a single model for all videos captured from different viewpoints and finds a shared high-level virtual path to map them on a single canonical view. Therefore, the model does not require any labels or viewpoint information in the learning phase. The network is learned only once using the spatiotemporal motion features of the gait sequences from several viewpoints, later it is used to construct the cross-view gait descriptors for the gallery and the probe sets. The descriptors are classified using simple linear support vector machine. Experiments carried out on the benchmark cross-view gait dataset, CASIA-B, and comparisons with the state-of-the-art demonstrate that the proposed method outperforms the existing cross-view gait recognition algorithms.

Gait has emerged as a distinguishable human biological trait. It refers to the walking style of an individual and is considered an important biometric feature for person identification. Codebook based gait recognition algorithms have demonstrated excellent performance by achieving high recognition rates. However, such methods construct a codebook for each database or scenario. In this paper, we investigate the idea of using a generic codebook for gait recognition. The proposed codebook is built by using spatiotemporal characteristics of gait sequences from a large diverse synthetic gait database. We also propose a gait recognition algorithm based on this generic codebook. The advantages of the proposed algorithm over the existing methods include its independency from generating a codebook for each database, rather the proposed generic codebook can be used to encode any gait scenario. Moreover, the proposed algorithm is model free and does not require human body segmentation or modeling. The performance of the proposed generic codebook-based gait recognition algorithm is evaluated on two large gait databases TUM GAID and CMU MoBo, and recognition rate reveals the effectiveness of the proposed algorithm.

Biometric gait recognition has received substantial attention of researchers in the recent years due to its applications in numerous fields of computer vision, particularly in visual surveillance and monitoring systems. Most existing gait recognition algorithms solve the problem of person identification either by constructing a human body model based on various skeletal data characteristics such as joints positioning and their orientation, or use gait features, e.g., stride length, gait patterns and other shape templates. Such approaches require the extraction of the human-body's silhouette, contour, or skeleton from the images, and therefore their performance highly depends on the silhouette segmentation accuracy. In this paper, we propose a novel gait recognition algorithm which exploits spatiotemporal motion characteristics of a person, which does not need silhouette or skeleton extraction at all. The proposed algorithm computes a set of spatiotemporal features from the video sequences and uses them to generate a codebook. Fisher vector is used to encode the motion descriptors which are classified using linear Support Vector Machine (SVM). The proposed algorithm is evaluated on three benchmark gait datasets: TUM GAID, CASIA-B, and CASIA-C. It achieved excellent results on all datasets which demonstrate the effectiveness of the proposed algorithm.

Multiview video plus depth (MVD) is the most popular 3D video format where the texture images contain the color information and the depth maps represent the geometry of the scene. The depth maps are exploited to obtain intermediate views to enable 3D-TV and free-viewpoint applications using the depth image based rendering (DIBR) techniques. DIBR is used to get an estimate of the intermediate views but has to cope with depth errors, occlusions, imprecise camera parameters, re-interpolation, to mention a few issues. Therefore, being able to evaluate the true perceptual quality of synthesized images is of paramount importance for a high quality 3D experience. In this paper, we present a novel algorithm to assess the quality of the synthesized images in the absence of the corresponding references. The algorithm uses the original views from which the virtual image is generated to estimate the distortion induced by the DIBR process. In particular, a block-based perceptual feature matching based on signal phase congruency metric is devised to estimate the synthesis distortion. The experiments worked out on standard DIBR synthesized database show that the proposed algorithm achieves high correlation with the subjective ratings and outperforms the existing 3D quality assessment algorithms.

In free-viewpoint television (FTV) framework, due to hardware and bandwidth constraints, only a limited number of viewpoints are generally captured, coded and transmitted; therefore, a large number of views needs to be synthesized at the receiver to grant a really immersive 3D experience. It is thus evident that the estimation of the quality of the synthesized views is of paramount importance. Moreover, quality assessment of the synthesized view is very challenging since the corresponding original views are generally not available either on the encoder (not captured) or the decoder side (not transmitted). To tackle the mentioned issues, this paper presents an algorithm to estimate the quality of the synthesized images in the absence of the corresponding reference images. The algorithm is based upon the cyclopean eye theory. The statistical characteristics of an estimated cyclopean image are compared with the synthesized image to measure its quality. The prediction accuracy and reliability of the proposed technique are tested on standard video dataset compressed with HEVC showing excellent correlation results with respect to state-of-the-art full reference image and video quality metrics.

Depth image based rendering is a well-known technology for the generation of virtual views in between a limited set of views acquired by a cameras array. Intermediate views are rendered by warping image pixels based on their depth. Nonetheless, depth maps are usually imperfect as they need to be estimated through stereo matching algorithms; moreover, for representation and transmission requirements depth values are obviously quantized. Such depth representation errors translate into a warping error when generating intermediate views thus impacting on the rendered image quality. We observe that depth errors turn to be very critical when they affect the object contours since in such a case they cause significant structural distortion in the warped objects. This paper presents an algorithm to improve the visual quality of the synthesized views by enforcing the shape of the edges in presence of erroneous depth estimates. We show that it is possible to significantly improve the visual quality of the interpolated view by enforcing prior knowledge on the admissible deformations of edges under projective transformation. Both visual and objective results show that the proposed approach is very effective.

The success of 3D and free-viewpoint television largely depends on the efficient representation and compression of 3D video in addition to viable rendering methods. This paper presents a novel 3D video coding technique based on the creation of a panorama view to compact the information of a stereoscopic pair. The panorama view represents the information that would be visible to a virtual camera with a larger field of view embracing all the available views. The information in the panorama view is then used to estimate any intermediate view using depth image based rendering. Furthermore, to fill the disocclusions in the reconstructed view a directional depth aided fast marching inpainting technique is presented. The panorama view and corresponding depth map are amenable to standard video compression. In this paper we show that using the novel HEVC standard the proposed 3D video format can be compressed very efficiently.

This paper presents a simple yet efficient image retrieval technique that defines image feature descriptors using localized multi-texton histogram. The proposed technique extracts a unique feature vector for each image in the image database based on its shape, texture and color. First, the image is divided into smaller equal size blocks and then for each block texture orientation is computed independently. Second, each block is filtered with a set of predefined textons and finally, a co-occurrence relation is established from the orientation and the filtered text on image. This relationship in turn provides a powerful feature vector. To retrieve similar images, the feature vector of the query image is computed and compared with the feature vectors of the stored images using Euclidean distance measure. The proposed algorithm is tested on standard image dataset Corel 1000 for accuracy and recall with favorable results. It is also compared with existing state of the art Context Based Image Retrieval algorithm and showed convincing results.

Three-dimensional video has gained much attention during the last decade due its vast applications in cinema, television, animation and virtual reality. The design of intermediate view synthesis algorithms that are efficient both in terms of computational complexity and visual quality is a paramount goal in the fields of 3D free view point television and displays. This papers focuses on the design of a low complexity view synthesis algorithm that produces better quality of the virtual image. A novel view synthesis technique to create a virtual view from two video sequences with corresponding depths is proposed. The technique employs low complexity integer pixel precision warping and a novel approach for hole filling based on inverse mapping. The proposed technique is tested over a number of video sequences and compared with existing state of the art methods, yielding excellent results both in terms of signal to noise ratio and visual quality.

Depth image based rendering of intermediate views with high visual quality remains a challenging goal in presence of estimated and quantized depth values. Among the other rendering artifacts we observed that edges are usually affected by significant warping errors. In particular, because of depth estimation inaccuracy around object boundaries the edges may completely loose their original shape during the warping process. Nonetheless, edges represent one of the most important cues for the human visual system. In this paper a novel technique aiming at improving the edge rendering is presented. As opposed to previous approaches, the technique exploits only texture information, thus avoiding possible errors in depth estimation. The idea is based on the enforcement of prior knowledge of the edge shape under projective transformation. The proposed algorithm works in two steps: first the damaged edges of the warped image are detected, then these latter are corrected so as to better approximate their shape in the reference view. Finally the corrected edges are rendered within the intermediate image without introducing noticeable texture artifacts. The proposed algorithm has been tested on a variety of standard video sequences exhibiting excellent results in terms of rendered image visual quality.

Most human detection algorithms in depth images perform well in detecting and tracking the movements of a single human object. However, their performance is rather poor when the person is occluded by other objects or when there are multiple humans present in the scene. In this paper, we propose a novel human detection technique which analyzes the edges in depth image to detect multiple people. The proposed technique detects a human head through a fast template matching algorithm and verifies it through a 3D model fitting technique. The entire human body is extracted from the image by using a simple segmentation scheme comprising a few morphological operators. Our experimental results on three large human detection datasets and the comparison with the state-of-the-art method showed an excellent performance achieving a detection rate of 94.53% with a small false alarm of 0.82%.