Narrow Results

Trajectory-clustering-based methods have shown a good performance in counting moving objects in densely crowded scenes. However, they still fall into trouble in complex scenes, such as with the close proximity of moving objects, freely moving parts of objects, and different object size in different locations of the scene. This paper proposes a new method combining velocity and location-specific spatial clues in trajectories to deal with these problems. We first extract the velocities of a trajectory over its life-time. To alleviate confusion around the boundary regions between close objects, extracted velocity information is utilized to eliminate unreal-world feature points on objects' boundaries. Then, a function is introduced to measure the similarity of the trajectories integrating both of the spatial and the velocity clues. This function is employed in the Mean-Shift clustering procedure to reduce the effect of freely moving parts of the objects. To address the problem of various object sizes in different regions of the scene, we suggest a technique to learn the location-specific size distribution of objects in different locations of a scene. The experimental results show that our proposed method achieves a good performance. Compared with other trajectory-clustering-based methods, it decreases the counting error rate by about 10%.

A computer vision method is presented for the mobile robot to find humans in scene. Face detection is used for confirming humans. In order to reduce regions of search, optical flow algorithm is used to segment the image in advance. Asymmetric problems in face detection are explained, and relative solutions are put forward by bootstrapping strategy and asymmetric adaboost algorithm. In addition, fisher discriminant analysis further improves the performance of face detection. Multi-view face models are trained to accommodate practical face detection application. At last, experiments demonstrate that our multi-view face detector achieves high detection accuracy and fast detection speed on both standard testing datasets and real-life images.

This paper proposes a new background subtraction method for detecting moving objects (foreground) from a time-varied background. While background subtraction has traditionally worked well for stationary backgrounds, for a non-stationary viewing sensor, motion compensation can be applied but is difficult to realize to sufficient pixel accuracy in practice, and the traditional background subtraction algorithm fails. The problem is further compounded when the moving target to be detected/tracked is small, since the pixel error in motion compensating the background will subsume the small target. A Spatial Distribution of Gaussians (SDG) model is proposed to deal with moving object detection under motion compensation that has been approximately carried out. The distribution of each background pixel is temporally and spatially modeled. Based on this statistical model, a pixel in the current frame is classified as belonging to the foreground or background. For this system to perform under lighting and environmental changes over an extended period of time, the background distribution must be updated with each incoming frame. A new background restoration and adaptation algorithm is developed for the time-varied background. Test cases involving the detection of small moving objects within a highly textured background and a pan-tilt tracking system based on a 2D background mosaic are demonstrated successfully.

Reformulating the Costeira–Kanade algorithm as a pure mathematical theorem, we present a robust segmentation procedure, which we call subspace separation, by incorporating model selection using the geometric AIC. We then study the problem of estimating the number of independent motions using model selection. Finally, we present criteria for evaluating the reliability of individual segmentation results. Again, model selection plays an important role. We confirm the effectiveness of our method by experiments using synthetic and real images.

In this work we present a framework for on-the-fly video transcoding that exploits computer vision-based techniques to adapt the Web access to the user requirements. The proposed transcoding approach aims at coping with both user bandwidth and resources capabilities, and with user interests in the video's content. We propose an object-based semantic transcoding that, according to the user-defined classes of relevance, applies different transcoding techniques to the objects segmented in a scene. Object extraction is provided by on-the-fly video processing, without manual annotation. Multiple transcoding policies are reviewed and a performance evaluation metric based on the Weighted Mean Square Error (and corresponding PSNR), that takes into account the perceptual user requirements by means of classes of relevance, is defined. Results are analyzed by varying transcoding techniques, bandwidth requirements and video types (with indoor and outdoor scenes), showing that the use of semantics can dramatically improve the bandwidth to distortion ratio.

This paper presents a vision based technique to identify the unspoken phones using a small camera that is located on the headset of the speaker. The system is based on temporal integration of the video data to generate motion history image (MHI). The paper proposes the use of global features to classify the MHI and compares the use of image moments with Discrete Cosine Transform (DCT). A comparison between Zernike moments (ZM) with DCT indicates that while the accuracy of classification for both techniques is very comparable (96% for ZM and 94% for DCT) when there is no relative motion between the camera and the mouth, ZM is resilient to rotation of the camera and continues to gives good results despite rotation but DCT is sensitive to rotation.

Based on the accuracy of the system and its resilience to movement artefacts such as rotation, the authors propose the use of such a system for human computer interface. Such a system could be invaluable when it is important to communicate without making a sound, such as giving passwords when in an open office or in public spaces.

A new approach of human activity monitoring with a single Inertial Measurement Unit (IMU) capable of gait recognition and assessment is proposed for home-based applications. The method estimates the foot motion using a single IMU, then automatically segments the motion into steps, and extracts multiple kinematics templates. It classifies each segment by extracting Mahalanobis distance-based features from multiple sections of the motion templates and then training a Support Vector Machine. The proposed wearable system can distinguish between nine classes of activities with a classification accuracy of 99.6%. It can also discriminate between normal and abnormal gait patterns with an accuracy of 98.7%. In addition to a high recognition rate, the proposed approach provides a Gait Similarity Score (GSS) of the performed gait to its desired/normal pattern. The experimental results indicate the capability of GSS measure for assessing the quality of motion in “pre-”, “initial”, “mid” and “terminal” stages of swing phase.

Control for and interaction with humanoid robots is often restrictive due to limitations of the robot platform and the high dimensionality of controlling systems with many degrees of freedom. We focus on the problem of providing a "skill-level interface" for a humanoid robot. Such an interface serves as (i) a modular foundation for structuring task-oriented control, (ii) a parsimonious abstraction of motor-level control (e.g. PD-servo control), and (iii) a means for grounding interactions between humans and robots through common skill vocabularies. Our approach to constructing skill-level interfaces is two-fold. First, we propose a representation for a skill-level interface as a "behavior vocabulary," a repertoire of modular exemplar-based memory models expressing kinematic motion. A module in such a vocabulary encodes a flow field (or gradient field) in joint angle space that describes the "flow" of kinematic motion for a particular skill-level behavior, enabling prediction from a given kinematic configuration. Second, we propose a data-driven method for deriving behavior vocabularies from time-series data of human motion using spatio-temporal dimension reduction and clustering. Results from evaluating an implementation of our methodology are presented along with the application of derived behavior vocabularies as predictors towards on-line humanoid trajectory formation and off-line motion synthesis.