Abstract:This paper presents the integrated service robot control platform developed at the Cognitive Technology Research Center in Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences. The platform is equiped with a sensor named Kinect. We can extract 2D and 3D features from the Kinect’s color images, depth images and point clouds, then classify the objects’ geometric models and choose the right grab attitude for the robot’s hand after those two types of features have been fused. Following the theory of Learning from demonstration (LfD), we have developed an approach to improve robot’s cognitive learning ability for indoor object manipulation tasks in everyday household environments, such as identifying the door handle’s location and opening the door, recognizing the targets in the cupboard, grabbing the object and sending it to the designated location, etc. Finally, we design an experiment to prove our method, which can teach the robot to grasp some cylindrical, rectangular or other geometrical objects and then finish the complicated task of planning the motion trajectory while interacting with the surrounding environment.

Abstract:Electromagnetic localization and orientation method has the advantages of high accuracy, high speed and high applicability, so that it is a reasonable choice to be used in surgical, indoor and outdoor robots. It uses alternating signals as excitation source. The generating coils (used as the excitation part) are applied with low-frequency AC currents to generate alternating electromagnetic field in space, while sensor coils (used as the sensing part) output the same frequency signal. By using the amplitude and phase information of the output signals from the sensor coils, the position and orientation information of the sensor coils related to generating coils can be calculated. In this paper, the electromagnetic localization and orientation system is presented toward the robot localization, including the magnetic model, localization algorithm and the hardware\software of the system. Two different excitation methods are used in the system, which are the time-sharing with 3-axis generating coils and phase-division with 2-axis generating coils. Sensor coils used in these two types are 3-axis coils. Experimental results show that the system can have a good accuracy with 1 mm.

Abstract:Road detection is of high importance in different advanced driver-assistance systems. It is widely used for functionalities such as pedestrian detection, obstacle avoidance, autonomous navigation, etc. The vision-based road detection is to classify image pixels belonging to road surface or not. Up to now, most road detection algorithms are designed for working during daytime. In this paper, we mainly focus on road detection at night. A near-infrared camera which provides infrared lamps to strengthen the weak illumination is used for image capturing. Firstly a planar reflection model is proposed to fit the intensity distribution of the images pixels. Next, a pixel-based classification is applied to determine whether the pixel is on the road surface or not. In the experiments, we compare our algorithm with the region growing method. The experiments show that our approach works better in some aspects.

Abstract:Phase singularities (PSs) refer to the zero points in complex signals. In previous studies, we found that PSs can provide rich information for image representation and are robust to image transformation, noise etc. This paper makes use of PS and bag of visual words (BoVW) model to construct bag of PS representation for images. Then we use SVM to classify bag of PS representations. Compared with previous works using SIFT points, our new representations not only use more interested points, but also allow us to pre-classify the words according to the sign property of PSs. The experimental results show that the proposed methods achieve better performance on PASCAL2005 image classification tasks than SIFT detectors.

Abstract:This paper presents a hand posture detection method based on transform feature representation and hierarchical model. The hierarchical model comprises a series of appearance models and an overall discriminate model. Appearance model for each posture is composed of a general template as well as several sub-category templates. With all the sub-category templates as transition functions, the original gradient histogram features can be converted into a more discriminative representation form. This transform representation is used to construct the discriminative model in the hierarchy model to achieve further posture-background and posture-posture classification. Moreover, to boost the efficiency, a skin-filter is introduced to exclude a wide range of non-skin area. Experimental results show that the proposed algorithm can successfully cope with appearance variability caused by viewpoint changes, posture tilts and natural posture deformation with a detection speed up to 20 frames per second.

Abstract:To solve the poor quality image transmission, this paper proposes a method to improve the transfer rate and reduce error rate of the image by using-MIMO (Multiple Input Multiple Output) technology for capsule endoscopy. In order not to increase the size and power consumption of the transmitting end, we use a single transmission antenna, while at the receiving end, we use multiple receiving antennas, that’s SIMO (Single Input Multiple Output). In the experiments, we simulated the performance of receive diversity techniques, and selected an optimal method of diversity combination.

Abstract:Transient elastography can be used to measure the viscoelastic properties of soft tissues using a simple inverse problem approach. However, there is measuring bias in this method due to several factors, e.g. radius of disk load and nearfield effect is the main reason for overestimating the shear velocity, and diffraction effect leads the measurement of shear attenuation to be overestimated. A self-adapting method is proposed to correct both the shear velocity and attenuation biases in our study. Theoretical and experimental results are in good agreement, and this new method is proved to be an effective way to correct the viscoelastic properties bias measured by transient elastography.

Abstract:The World Health Organization found that cardiovascular and heart disease causes the highest probability of death in the world. Electrocardiogram (ECG) is an important tool widely used in clinical prevention and diagnosis of cardiovascular and heart disease. Automatic analysis of ECG diagnostic technique can greatly reduce the workload of the cardiologists and improve the efficiency of diagnosis. The classification of ECG Heartbeat is the mainly research direction of ECG automatic analysis as automatic ECG heartbeat classification can improve the diagnostic quality of arrhythmia, especially in the area of dynamic electrocardiogram or the long-term ECG recording. This paper presents an ECG beat classification algorithm, The algorithm uses clusting analysis, mixed with linear classifiers, weighted judgment and physician-assisted classification. Using MIT-BIH-AR arrhythmia database as the raw data and ANSI / AAMI EC57: 1998 / (R) 2003 of AAMI as the standard of classification, the experiment results show that only using clusting analysis, mixed with linear classifiers and weighted judgment, the accuracy rate is 86.60%. After introducing cardiologists-assisted classification, the final accuracy rate is 98.16%.

Abstract:Human soft tissues generally exhibit complex material perperties such as nonlinearity, anisotropy, incompressibility and viscoelastictity. Soft tissue deformation is one of the most important yet difficult research tasks in virtual surgery. This paper presents a comprehensive survey on simulation of soft tissue deformation in virtual surgery. We first give an introduction of the virtual surgery system. Then we detailed various methods from geometrically-based methods to physically-based methods, from mesh-based models to meshless models. Finally, we describe some promising research directions on this topic.

Abstract:This resource deployment strategy for multi-user cluster gives priority to the fair allocation of resources between different users, while achieving resource sharing. In order to make full use of cluster resources, certain principles are used as weights for sharing cluster resources. It cares about whether the resources ratio which each user gets is in line with the principle, but not the total amount of resources, thus can ensure that the idle resources are utilized fully. The strategy consists of fixed allocation and independent application to meet users’ need. Then we propose the allocating pattern based on resources and taken time together to adapt to the more complex situation.

Abstract:The Rapid growth of data has provided us with more information, yet challenges the traditional techniques to extract the useful knowledge. In this paper, MCMM, a Minimum spanning tree (MST) based Classification model for Massive data with MapReduce implementation is proposed. It can be viewed as an intermediate model between the traditional K nearest neighbor method and cluster based classification method, aiming to overcome their disadvantages and cope with large amount of data. In this model, we treat the training set as weighted undirected complete graph. The vertices are objects and the weight of an edge between two objects is their distance, which could be a certain distance metric other than Euclidean distance. Then we find a minimum spanning tree forest of the graph, in which each tree represents a class. In order to reduce the computing time, we extract the most representative points of each tree to represent that tree. The shrunk point sets can be used for classification by computing the distances from unlabeled objects to them.MCMM model is implemented on Hadoop platform, using its MapReduce programming framework. Since Hadoop supports data intensive distributed applications and enables applications to work with thousands of nodes and petabytes of data, MCMM model is scalable to deal with massive data. In addition, MapReduce and Hadoop work well on cluster composed of commodity machines. Therefore there is no special need for particular hardware or architecture. This is actually the feature of cloud computing. MCMM model is used on cloud platform and could benefit from cloud computing by using Hadoop and MapReduce. Experiments had been carried out on several data sets including real world data from UCI repository and synthetic data, using Downing 4000 cluster, installed with Hadoop. The results show that MCMM model outperforms KNN and some other classification methods on a general basis with respect to accuracy and scalability.