Search

International audience; This paper tackles the problem of discovering subtle fall risks using skeleton clustering by multi-robot monitoring. We aim to identify whether a gait has fall risks and obtain useful information in inspecting fall risks. We employ clustering of walking postures and propose a similarity of two datasets with respect to the clusters. When a gait has fall risks, the similarity between the gait which is being observed and a normal gait which was monitored in advance exhibits a low value. In subtle fall risk discovery, unsafe skeletons, postures in which fall risks appear slightly as instabilities, are similar to safe skeletons and this fact causes the difficulty in clustering. To circumvent this difficulty, we propose two instability features, the horizontal deviation of the upper and lower bodies and the curvature of the back, which are sensitive to instabilities and a data preprocessing method which increases the ability to discriminate safe and unsafe skeletons. To evaluate our method, we prepare seven kinds of gait datasets of four persons. To identify whether a gait has fall risks, the first and second experiments use normal gait datasets of the same person and another person, respectively. The third experiments consider that how many skeletons are necessary to identify whether a gait has fall risks and then we inspect the obtained clusters. In clustering more than 500 skeletons, the combination of the proposed features and our preprocessing method discriminates gaits with fall risks and without fall risks and gathers unsafe skeletons into a few clusters.

This paper presents 3D model generation of black cattle using multiple RGB cameras for their BCS. The use of advanced ICT has a certain possibility to improve various agricultural activities. The authors have such a project whose targets are beef cattle. The goal of the project is to capture 3D shape information of black cattle for the estimation of their body condition scores (BCS). Cattle are always moving because they are animals. Therefore, it is very difficult to capture their body shape information even using a commercial 3D scanner. Another reason is that the color of beef cattle is almost black and then a commercial 3D scanner like a laser range finder cannot be used. So, as the first trial, the authors used multiple RGB cameras to capture RGB images of a cow, generated manually its silhouette images, and employed Shape-from-Silhouette(SfS) method to generate its 3D model. The authors took multiple RGB camera images of cows in a natural environment and generated their 3D models. From the generated 3D models of cows, it can be found that it is possible to estimate the weight of each cow correctly if its accurate silhouette images are generated manually. Here, the problem is how the accurate silhouette images can be obtained automatically in a natural environment. From several experiments, the authors conclude it is impossible. Therefore, the authors propose the use of new method based on multicolor attributed voxels instead of SfS method. This paper clarifies the availability of the new method by showing several experimental results.

This paper presents 3D model generation of cattle by shape-from-silhouette method for ICT agriculture. The use of advanced ICT has any possibility to improve various agricultural activities. The authors have such a project whose targets are beef cattle. The goal of the project is to capture 3D shape information of cattle for the estimation of their body condition scores (BCS). Cattle do not stop moving because they are animals. Therefore, it is very difficult to capture their body shape information even using a commercial 3D scanner. Another reason is that the color of beef cattle is almost black and then a commercial 3D scanner like a laser range finder cannot be used. The authors use multiple RGB cameras to capture silhouette images of a cow and employ shape-from-silhouette method to generate its 3D model. Actually, the authors have taken multiple RGB camera images of cows and generated their 3D models. And then, it can be found that the generated 3D models' volumes of cows have positive correlation with their weights. This result says that the estimation of cows' weights is possible from multiple RGB camera images of them.

Understanding the states of learners at a lecture is useful for improving the quality of the lecture. A video camera with an infrared sensor Kinect has been widely studied and proved to be useful for some kinds of activity recognition. However, learners in a lecture usually do not act with large moving. This paper evaluates Kinect for use of activity recognition of learners. The authors considered four activities for detecting states of a learner, and collected the data with the activities by a Kinect. They applied K-nearest neighbor algorithm to the collected data and obtained the accuracy 0.936 of the activity recognition. The result shows that Kinect is applicable also to the activity recognition of learners in a lecture.

International audience; This paper proposes a novel method to detect objects by a mobile robot which adapts to an environment. Such a robot would help human designers of a smart environment to recognize objects in the environment with their attributes, which signi cantly facilitates his/her design. We rst introduce Lifting Complex Wavelet Transform (LCWT) which plays an important role in this work. Since the LCWT has a set of controllable free parameters, we can design the LCWTs with various properties by tuning their parameters. In this paper we construct a set of LCWTs so that they can extract local features from an image by multi-scale. The extracted local features must be robust against several kinds of changes of the image such as shift, scale and rotation. Our method can design these LCWTs by selecting their parameters so that the mobile robot adapts to the environment. Applying the new set of LCWTs to the images captured by the mobile robot in the environment, a local feature database can be constructed. By using this database, we implement an object detection system based on LCWTs on the mobile robot. Effectiveness of our method is demonstrated by several test results using the mobile robot.

This paper aims to develop a fast algorithm for detecting a new object from video sequences captured by on-board robot vision. We first propose lifting complex wavelet, which is a new method for extracting local features in an image. The proposed lifting complex wavelet transforms can be detected the features faster than the conventional SIFT algorithm. Our new object detection is performed by using an overlap ratio between the local features of current- and past-frames. In experiments, we show that new objects can be detected fast from on-board robot vision.

This paper proposes a motion retrieval system using Interactive Evolutionary Computation (IEC) based on Genetic Algorithm (GA) and motion features defined based on Laban Movement Analysis (LMA) used for the similarity calculation of motions in the system. The proposed IEC-based motion retrieval system allows the user to retrieve motions similar to his/her required motions easily and intuitively only through the evaluation repeatedly performed by scoring satisfaction points to retrieved motions without entering any search queries. The authors newly define LMA-based motion features to represent them as genes of GA used for the similarity calculation in the system. This paper also clarify that the LMA-based motion features are available as similarity features of motions by showing results of analyzing them using SOM visualization.

This paper proposes new motion generation method by Interactive Evolutionary Computation based on Genetic Algorithm. This method generates new motions by combining some primitive motions, which are obtained by dividing already existing motions. This combining process is performed by IEC based on GA and then combined motions are generated according to the user preference. There is a problem that the variation of generated motions is very small if the number of already existing motions is too small. To compensate this problem, the authors also employ motion signal processing technique. Generated motions by combining primitive motions are modified through multi-level decomposition and composition of discrete wavelettransform. When applying the composition process, each level wavelet coefficients are modified by their gain parameter which the user chooses through IEC and then motions modified by the user preference are generated. In this way, combinational use of IEC based on GA and motion signal processing technique allows the user to generate his/her required motions easily and intuitively.

This paper presents an interface of image query system for content-based image retrieval.The proposed interface can split images into the regions and select an target region from the separated regions. We apply content-based image retrieval to the selected region as a query.The images in databases are also split into the areas in advance.The content features such as color and texture are extracted from these areas and stored in the feature database.The texture features of images are based on statistic obtained by wavelet transform of images.Finally, our system presents the images whose areas are similar to the query's areas specified by our image query interface.In addition, our interface can also tune the weight parameters about the content features in order to get the desired images manually.

A person identification system based on fast face learning of lifting wavelet filters is proposed. The real power of our system lies in fast learning of lifting wavelet filters adaptive to facial parts such as eyes, nose and lips, in a set of training faces. In our system, free parameters in the lifting filter are learned fast by using Newton’s method. The learned parameters are memorized in a database together with the training faces. The lifting filters with the learned parameters in the database are applied to each of video frames which contain faces of a person, and the faces are detected by measuring some kind of distance. A person whose face is detected in a maximum number of frames is identified as a target person. To realize fast face detection, the learned filters are applied only to the skin areas separated from background by using color segmentation. Simulation results show that our person identification algorithm is accurate and fast.

This paper proposes a video scene search system with a sketch query interface, whose search algorithm is based on a stochastic ARG (Attributed Relational Graph) matching. In this system, using the sketch query interface, the user can enter a sketch image as his/her query intuitively for searching scenes of a video. As its preprocessing, the system divides a video into several different scenes by extracting scene-change (cut-change) frames using a kind of chi-square test between color distributions of two adjoining frame images of the video. For searching the user's required scenes similar to his/her entered sketch query image, the system decomposes each of the scene-change frame images into some layers and an ARG is constructed from the layers. The ARG is used as the feature of the scene for the similarity search. Each sketch image entered as a query is also decomposed into layers and represented as an ARG. Hence, the system finds scenes of the video similar to the user's sketch image entered as his/her query by a stochastic ARG matching algorithm.

This work presents a fast object-tracking algorithm using the lifting wavelet filters. The algorithm solves a linear system of simultaneous equations of free parameters in the lifting filter and searches a small region for objects. In simulation, it is shown that the algorithm can track the object in a sequence of images.

This work presents a personal identification system based on the learning of the lifting dyadic wavelet filters. Our system consists of face learning, detection, and identification processes. In the learning process, free parameters in the lifting filters are determined so as to capture a facial part. Our face detection method is performed by applying the learned filters to each of the video frames. A person whose face is detected in a maximum number of frames is identified as a target person. In simulation, it is shown that our personal identification algorithm is fast and accurate.

This paper presents a fast algorithm for detecting facial parts such as nose, eyes and lips in an image by using lifting dyadic wavelet filters. Free parameters in the lifting filters are learned so as to maximize the cosine of an angle between a vector whose components are the lifting filters and a vector of pixels in the facial part. Applying the learned filter to a test image, facial parts in the image is detected. In simulation, we show that our algorithm is fast and robust one for detecting facial parts from an image.

This paper proposes a method for extracting sub-images from a huge reference image by using learnt lifting wavelet filters. Lifting wavelet filters are biorthogonal wavelet filters containing free parameters developed by Sweldens. Our method is to learn such free parameters using some training sub-images so as to eliminate their high frequency components. The learnt wavelet filters have the feature of training sub-images. Applying such wavelet filters to the reference image, we can detect the area which has almost the same high frequency components as those of the target sub-image.

This paper describes a new method for finding portions having the same feature in target signals or images from a time series or a reference image. The new method uses an integer-type Haar lifting wavelet transform. Free parameters contained in this transform are learned by using training signals or images. The advantage of this method is to be able to find portions having the same feature in the targets, and to realize robust extraction due to rounding-off arithmetic in the trained transform. In simulations, we show how well the method finds geomagnetic sudden commencements from time series of geomagnetic horizontal components, and extracts facial images from a snapshot.

Recognition of moving objects is one of the most important problems in computer vision. This problem has many applications such as individual recognition, gesture recognition and lip reading. Murase and Sakai [1] proposed a method using parametric eigenspace representation to realize efficient recognition of moving objects. This method is independent of the applications and requires a little time to recognize the moving objects. However, it is needed to normalize the position and size of silhouettes extracted from moving images