Abstract:Collision detection is a key step in 3D virtual clothing, and it is difficult to be made real-time using generic collision detection algorithms with relatively high resolution of cloth and human body. This paper exploits depth and normal maps to efficiently detect and resolve collisions between cloth and body. First, body mesh is rendered from prepositioned cameras to generate depth and normal maps. Secondly, the depth of cloth node is computed and then the node is transformed to depth image space. Thirdly, depth is retrieved from the depth map according to the image coordinate of the cloth node and it is compared with the node's depth to determine whether collision happens or not. Lastly, if collision happens, an interpolation parameter is determined by searching the depth image space in the coordinate interval determined by the cloth node positions in previous and current integration step using a modified DDA line rasterization algorithm, and the interpolation parameter is later used to compute the contact point and contact normal which is necessarily for continuous collision response. Experimental results show that the algorithm takes little time in the preprocessing step and is able to provide real-time collision detection and response even when the resolution of cloth mesh and human body is relatively high.

Abstract:This paper addresses the problem of node dynamic selection in camera networks. A selection method based on reinforcement learning is proposed in which the node is selected to maximize the expected reward while minimizing the switching with Q-learning. To accelerate the convergence of Q-learning, the geometry of camera networks is considered for initial Q-values and a Gibbs distribution is used for exploitation. In order to evaluate visual information of the video, a function of the visibility, orientation, definition and switching is designed to assess the immediate reward in Q-learning. Experiments show that the proposed visual evaluation criteria can capture the motion state of the object effectively and the selection method is more accurate on reducing cameras switching compared with the state-of-the art methods.

Abstract:With the development of machine learning theory and image processing technology, peoples are more and more interested in how to build a system to automatically evaluate and assess aesthetics quality of photos in the field of computer vision and computational aesthetics. This type of system can be a supplement for the subjective assessment of photo aesthetic quality. In this paper, 25 visual features extracted from each image are used to objectively evaluate photo aesthetic quality, which can better reflect the aesthetic quality of portrait photo. Four aesthetics classifiers are built based on support vector machine, AdaBoost, random forest and linear regression. 10-Fold cross validation experiment is performed to reveal which features have a salient impact on the aesthetic assessment. Compared to the current research results, classifiers using 25 features proposed by this study have higher classification accuracy rate for portrait photo aesthetic evaluation, even using the smaller training sets.

Abstract:This paper presents an approach to reconstruct and optimize static 3D human body models using a single Kinect-like motion depth camera by moving the sensor freely around the human body. First, to reduce the noise in the depth data captured by the Kinect-like depth sensor, an approach is proposed to filter it according to the noise source created from the Kinect-like depth sensor. Then, the search for the corresponding patch is performed by combining with the information of depth and RGB, and the pair frames are aligned with the Gaussian mixture model and the improved signed distance function. For the loop closure problem, a closed curve fitting based method is provided with different layers to globally register all the frames. Finally, the human body surface is constructed by using the Poisson reconstruction. Experimental results show that the presented approach can reconstruct human models with high quality.

Abstract:Tetrahedralization of 3D models is an important technique in finite element mesh generation. An algorithm of tetrahedral mesh generation and optimization is proposed in this paper. It first uses principal component analysis of 3D models for the preprocessing. Then the initial tetrahedral mesh is built based on the body-centered cubic while the movement of cutting point is improved via Laplacian coordinates for keeping local features. Finally, it optimizes the quality of the tetrahedral mesh by using the function of energy-density error. Experimental results demonstrate that this method is viable, effective, and can preserve the model feature well.

Abstract:This paper proposes a price-based incentive mechanism for P2P anonymous communication system. Through pricing on dummy traffic, relay traffic and export traffic, the P2P anonymous communication system builds a pricing system. The pricing system quantifies the user's system resource production and consumption. First, it can effectively stimulate users in P2P anonymous communication system to provide relay and export services. These services effectively improve system performance. Secondly, the pricing system can motivate free-riders in providing dummy traffic. The dummy traffic enhances the system's anonymity. Finally, the pricing system is also capable of prompting the user to request the appropriate relay nodes according to their needs in order to avoid unnecessary consumption of system resources. User policy analysis based on scenarios confirms the validity of price incentives P2P anonymous communication system.

Abstract:In modern data centers, due to the deadline-agnostic congestion control in transmission control protocol (TCP), many deadline-sensitive flows cannot finish sending before their deadline. Therefore, providing high deadline meeting ratio becomes a critical challenge in typical OLDI (online data intensive) applications of DCN. Under the partition-aggregate workflow pattern, since all the responses of single request have the same flow deadline and size, they are likely to respond nearly at the same time, which may result in that all flows with similar priority miss their deadlines. In this paper, two kinds of deadline-aware TCP, HPD and P²D are proposed. By using the novel deceleration, HPD and P²D alleviate the impact of priority synchronization problem. At-scale NS2 simulations show that P²D reduces the fraction of missed deadlines by 20% compared to D²TCP.

Abstract:Successive interference cancellation (SIC) is an effective multipacket reception scheme to combat interference at the physical layer. This paper studies the problem of maximizing the number of successful simultaneous transmissions (i.e., transmission capacity) in a wireless network with SIC at the physical layer. First, an interference model based on the physical interference model is proposed to characterize the sequential detection nature of SIC. Then, an algorithm is presented to evaluate whether or not a link set is feasible. Next, recognizing capacity maximization is a NP-hard problem, a novel approximation solution based on the genetic algorithm (GA) is provided. Finally, the design and parameter setting of the GA solution are discussed, and the performance is validated by various simulations.

Abstract:Cloud computing has become the focus information processing and many related fields with its powerful computing and storage capacity. For some of the existing phenomenon about the large calculation and high computing time cost in cloud task scheduling algorithms based on the batch model and evolution algorithm, this paper presents a local cloud task scheduling algorithm based on improved GEP and change of resources. In the process of designing the algorithm, it is first improved the GEP algorithm according to the characteristics of cloud task scheduling. And then, this paper constructs a fitness function considering both the comprehensive utilization and energy consumption. Finally, this paper constructs a local cloud task scheduling algorithm based on improved GEP and comprehensive utilization. The algorithm proposed in this paper reduces the computing time cost by monitoring the physical resource usage and reducing the number of physical machines involved in the task scheduling. The comparison experiments among GEP, genetic algorithm and the algorithm proposed in this paper based on RH (rolling horizon) model has been made. The results show that the proposed algorithm can not only reduce the optimization time, hard to fall into the local optimal solution, but also has the faster convergence speed.

Abstract:Reduction of quality (RoQ) attack is an atypical denial of service (DoS) attack, which has a strong concealment. Consequently, most traditional methods of detection are no longer applicable. There are a number of new methods developed recently. However, most of these methods have higher false positive rate in varying degree. In this paper, a novel method is proposed based on the principle of time-frequency analysis with Wavelet multi-resolution and Cepstral technique. First, according to different time-domain characteristics, the potential anomaly is detected and the abrupt change point is located. Secondly, the local traffic around the abrupt change point is analyzed by cepstrum. The potential characteristics of attack periodicity is extracted. By the two-stage detection, this new method ultimately can confirm whether the network is affected by the attack. Results of simulations and real network experiments demonstrate that the presented algorithm can detect RoQ attacks accurately with very low false positive rate and false negative rate.

Abstract:The dense deployment of wireless access points (APs) makes AP association an important problem. Currently, AP association is solely based on the signal levels of APs. However, this approach fails to consider the heterogeneous nature of APs, and the variety of user demands (bandwidth, security, delay, etc). In addition, distributed AP association cannot achieve network-level load balance. To address the issue, this paper proposes a centralized AP association model based on the software defined network (SDN). This model considers the objective of network administrator and wireless clients simultaneously, and can optimize the load balance of APs as well as satisfactory of clients. Ant colony algorithm is used to solve the model, and simulation is performed to validate the algorithm. Results show that user satisfactory factor as defined in our model increases from 54.5% to 86.8% under heavy load, and to 94.1% under light load. In addition, the load balance of APs also improves remarkably.

Abstract:With the popularity of mobile devices and social informatization, a lot of trajectory data is increasingly accumulated in daily life and used in different applications. Based on the location information of users in the mobile social network and campus informatization management network, a trajectory tracking system named Argo is proposed in this paper. In order to realize multi-information fusion trajectory tracking, various sources of location information such as Weibo, E-mail, BBS, E-card and the passive localization information from wireless access points covering the entire campus, are all analyzed in the Argo. Experimental results show that the system can effectively achieve the trajectory tracking and provide a better service for users.

Abstract:The rapid spreading of mobile devices and smartphones has promoted the deployment of enterprise WLANs. This research collects read traces from over 10000 WiFi users in one university in Shanghai. Through intensive data analysis, it finds that there exists more than 24% redundancy traffic in WLANs. By mining the patterns of data set, a user-oriented redundancy elimination strategy is proposed to allocate different cache sizes to different users adaptively. Simulation results show that user-oriented redundancy elimination strategy can detect and eliminate redundant traffic effectively.

Abstract:In computer vision and multimedia areas, it's an important yet challenging problem to perceive human motion at semantic level. In this work, a novel approach is presented to map the low-level response to semantic description of human actions. The features are based on the detection of deformable part models, in which the body pose information is contained implicitly under the specific human actions. The filter responses of the detectors are mapped to an effective feature description, which encodes the position and appearance information of human body and parts. The obtained features capture the relative configuration of body parts, and are robust to the false detections occurred in the individual part detectors. Comprehensive experiments conducted on three databases show the presented method achieves remarkable performance in most of the cases.

Abstract:Extracting nodes that reflect image content and assigning initial labels for these nodes are two critical technologies for saliency detection. A novel method of saliency detection is proposed by this work. It consists of two main parts, one is self organizing map (SOM), and the other is manifold learning (ML). Hundreds of nodes are obtained by the SOM. These nodes can capture not only the color, but also the contour of image content. By means of embedding a two dimension map into higher Euclid space, a weighted undirected graph is constructed. In consideration of edge symmetry in undirected graph, a manifold learning method, which combines undirected graph and semi-supervision, is further proposed. With supplied initial saliency values for nodes along image borders, the saliency values are computed for all nodes. Experimental results demonstrate the proposed model not only achieves high performance on precision and recall, but also presents a pleasing visual effect.

Abstract:Inferring user attributes is important for user profiling, retrieval, and personalization. Most existing work infers user attribute independently and ignores the relations between attributes. In this work, a new method is proposed to infer user attributes via hypergraph learning. In the hypergragh, each vertex represents a user in the social media, and the hyperedges are used to capture the similarity relations of the user generated content and the relations between attributes. The user attributes inference is formalized into a regularization label similar propagation problem in the constructed hypergraph, which can effectively infer the users' various attributes. Extensive experiments conducted on a collected dataset from Google+ with full attribute annotations demonstrate the effectiveness of the proposed approach in user attribute inference.

Abstract:Action and behavior analysis of players is a direct method of high-level semantic analysis or highlight annotation in sports video. Accurate detection and segmentation of players is the key technology of this method. Employing domain knowledge and characteristics of mid-level feature patch in sports video, a semi-supervised algorithm is proposed to discover the mid-level feature patch and train the player detector for different types of video shots. The detection result is used to label the superpixel, and then player segmentation is accomplished by Grab Cut segmentation algorithm. Experimental results show that the mid-level feature patch based player detector is convenient to train and achieves high detection accuracy. The detected player regions can be used to segment the players effectively, and hence the computation procedure of player segmentation is simplified.

Abstract:With the fast development of mobile devices and great concern on health surveillance, it is becoming increasingly popular to collect, analyze, and interpret people's health-related data by using intelligent mobile devices in their daily life. Oxygen saturation is an important physiological parameter referring to the concentration of oxygen in the blood, and prolonged low oxygen levels may lead to respiratory or cardiac arrest. Previous oxygen saturation detection methods require infrared light, however most of the off-the-shelf mobile devices lack such infrared light transmitter and receiver modules. This paper presents a novel oxygen saturation estimation method emplying a RGB camera and visible light in most of mobile devices. By applying of traditional optical oxygen saturation estimation model to mobile devices, and analyzing its problem for those devices, this study proposes a new oxygen saturation estimation model on intelligent mobile devices and offers an approach to solve the baseline drift problem in mobile cameras. Experiments demonstrate the effectiveness of the proposed method and its potentials in many oxygen saturation based researches such as daily-activity based healthcare with mobile devices.

Abstract:Natural multimodal human computer interaction dialog requires computer be able to produce intelligent response to user's statement. Due to the limitations of knowledge base and randomness of user's discourse, a traditional human-computer dialogue system cannot answer or produce consistent answer with user's expectations when the conversation is beyond the scope of knowledge, thus affecting user's sense of experience to the natural machine dialogue system. To solve this problem, this paper presents a method of generating optimal sentence by integrating multi-modal interaction history information and data-oriented parsing model. First, rules of context-free grammar from large-scale syntax tree libraries are extracted. Then combining user's expressions, gestures and other multi-modal interaction history information in dialogue process, a data-oriented parsing (DOP) model is integrated to filter Chinese sentences which are generated by context-free grammars, ultimately generating a sentence which is grammatically and semantically sound. The method allows a computer to generate responses to the current dialogue according to the interaction history information when the system can't get the support of knowledge base, therefore enhancing user's experience to multi-channel natural-machine interaction system. The proposed method is applied to traffic information search and multi-modal multi-topic dialogue system, and the result shows it can effectively improve the naturalness and enhance user's experience.

Abstract:Understanding and optimizing the power consumption of smartphones has become an important research topic. It is necessary to have a dynamic power estimation tool for hardware and software developers so that they can develop energy-efficient applications and construct energy-efficient smartphone systems. Previous work has proposed various power models for estimating the power consumption. However, these models lack granularity and accuracy. In this paper, a power model for smartphones considering hardware utilization and power delay is proposed. The model makes each hardware component more fine-grained and includes the power delay. Therefore, it can more accurately estimate the real-time power consumption. The model is based on a nonlinear regression structure. First, the model is determined by making each system variable modular from the target device. Then, the specified model is identified by test cases and the final coefficients concerning the power consumption are confirmed. Finally, the estimated power is compared with the actual power measured. Experimental results demonstrate that the average absolute error of power model is less than 4.6% in common scenarios, which obviously improves the accuracy of evaluation.

Abstract:Using Kinect camera combined with augmented reality and hand gesture recognition method, this paper designs and implements a virtual playing system for bowstring musical instruments (e.g. erhu). The fusion rendering of real scene captured by Kinect and virtual instrument forms the augmented reality scene. Through the depth data and a Bayesian skin model, the system segments the user's left hand region, then draws the hand image on the enhanced image, thus solving the problem of occlusion between virtual and real worlds in augmented reality system. Based on inverse kinematic methods and Markov model the presented design also provides a 3D virtual gesture fitting method to recognize left hand gestures, then completes the virtual instruments playing combined with right hand movement status.

Abstract:Color contrast is an important cue for image attention region detection. Extracting image regions that contain distinguishing color features is very helpful for computing the contrast of each image region. To obtain an efficient contrast map, the closure prior is firstly exploited to pick up the image regions containing distinguishing color features via connectivity detection in layered bit-planes. Secondly, the background prior is used to remove closed regions that touch image boundaries, and obtain closed region masks, in which the elements of closed regions are labeled with "1". Thirdly, a hypothesis, that a region should have big chance to be an attention region if it appears more times as a closed region in layered bit-planes, is proposed based on the contrast and closure priors. Further, the closed region masks of all bit-planes are accumulated to obtain the contrast of each connected region. Meanwhile, by taking account of the characteristics of human visual system with respect to the perception for small attention region, and visual resource allocation, several morphological filtering technologies are adopted to the key steps of contrast computing. Finally, the saliency map oriented to visual fixation estimation is generated. The experimental results show the presented detection method achieves acceptable performance compared with several state-of-the-art models.

Abstract:Query routing among peers to locate resources is a main issue discussed in Peer-to-Peer (P2P) networks, especially in unstructured P2P networks. This issue becomes even worse when frequent join and departure or failure of client peers happen in the networks. This paper proposes a new churn-resilient protocol to assure alternating routing path to balance query among peers under network churn. The proposed protocol uses two strategies to make queries balancing among inter-and intra-group peers. First a resource grouping and rewiring strategy is provided to periodically cluster peers having same set of resources. This grouping strategy makes locating resources more efficient in the inter-group peers, for it makes the network overlay topology evolve from a random network to a clustered network. The rewiring strategy also alleviates loads among over-loaded peers. Meanwhile, load balancing routing among the intragroup peers is achieved by collaborative Q-learning among peers. The collaborative Q-learning method not only learns from such parameters as processing capacity, number of connections and number of resources in peers, but also learns their state of congestion. Using this technique, queries are guided to avoid forwarding to those congested peers. Thus, query routing forwarding is balanced among intra-group peers. Simulation results show that the desired resources are located more quickly and queries in the whole network are balanced. The results also show that queries by the proposed method exhibit more robustness and adaptability under network attacking, high query workloads, and high network churns than queries by random walk method.

Abstract:A scalable locality sensitive hashing (SLSH) scheme is proposed to solve the problem of indexing high-dimensional data for dynamic datasets. The dynamic property destabilizes the size of the dataset, fuzzes up the tendency of data distribution, and conduces to the retrogression of retrieval performance. SLSH inherits two very convenient properties from the novel E2LSH that SLSH can rapidly work on data that is extremely high-dimensional and directly works on Euclidean space. For the purpose of adaptively fit the dynamic data distribution, the original hash family in E2LSH is altered for SLSH. A constraint of hash bucket capacity is applied for the hash parameters adjustment. As a result, SLSH provides robust partitions in the high-dimensional space for the dynamic data.

Abstract:This paper provides a comprehensive optimization strategy aiming at reducing the complexity of high efficiency video coding (HEVC) encoder with CPU-GPU cooperation. Based on the computational complexity distribution of HEVC encoder and characteristics of different modules and coding tools, intra coding, inter coding and in-loop filtering are collaboratively optimized. For intra coding, based on the correlation between neighboring coding units (CUs), depth range of CU is predicted and the number of candidates in intra mode candidate set for RDO (rate distortion optimization) is cut down, to avoid unnecessary computations. For inter coding, the most time consuming module, motion estimation (ME), is implemented with the collaboration of CPU and GPU in pipeline. Based on the energy of prediction residuals, an early termination scheme of CU splitting is proposed in this paper. For in-loop filtering, GPU based sample adaptive offset (SAO) parameter decision scheme and GPU based deblocking scheme are proposed to further reduce the coding complexity on CPU. The overall optimization scheme is implemented on the HM 16.2 platform, and experiments demonstrate that the proposed optimization scheme can reduce over 68% of the coding complexity of HEVC encoder, with only 0.5% performance loss in average.

Abstract:Maximum noise fraction (MNF) rotation is a classical method of hyperspectral image dimensionality reduction, and it needs a large amount of calculation and thus is time-consuming. This paper investigates the code transplantation and performance optimization for the maximum noise fraction algorithm on multi-core CPU and many integrated core (MIC) architecture. By analyzing hotspots of the MNF algorithm, parallel schemes are first designed for filtering, covariance matrix calculating and MNF transforming. Then, a series of optimization methods are presented and validated for various parallel schemes of different hotspots, including using math kernel library (MKL) functions. Finally, a C-MNF algorithm on multi-cores CPUs and an M-MNF algorithm on the CPU/MIC heterogeneous system are constructed. Experiments show that the C-MNF algorithm achieves impressive speedups (ranging from 58.9 to 106.4), and the M-MNF parallel algorithm runs the fastest, reaching a maximum speed-up of 137X.

Abstract:In mixed criticality systems, tasks with different criticality levels share a common platform, which makes the schedulability more complex. Considering AMC (adaptive mixed criticality) scheduling is currently the most effective fixed priority approach for scheduling mixed criticality systems, this work presents a response time analysis algorithm AMC-PM (AMC partition max) for AMC. In AMC-PM, the WCET (worst case execution time) of the task is partitioned into low critical execution time and high critical execution time. Then an upper bound of response time can be derived by adding the response times of the two parts together. For tasks with small WCET, evaluations illustrate that AMC-PM can significantly enhance the schedulability comparing with AMC-rtb and that AMC-PM can effectively decrease the run time comparing with AMC-max.

Abstract:With the expansion of the market share occupied by the Android platform, security issues (especially application security) have become attention focus of researchers. In fact, the existing methods lack the capabilities to manage application permissions without root privilege. This study proposes a dynamic management mechanism of Android application permissions based on security policies. The paper first describes the permissions by security policies, then implementes permission checking code and request evaluation algorithm in Android framework layer. Experimental results indicate that the presented approach succeeds in permission management of Android applications, and its system overhead is low, which makes it an effective method for Android permission management.

Abstract:Mobile software is often deployed on processors with limited energy. Energy consumption has been an important quality attribute to evaluate such software. Compared with the code level and instruction level assessment, energy consumption assessment at design level has the advantages of low time consumption and low cost. In recent years, it has become a research focus in academic and industrial fields of software engineering. Currently, most of the methods on energy consumption evaluation did not evaluate the energy consumption of internal behavioral elements of software components, resulting in the problem of low precision. To tackle this problem, this study builds a model of energy consumption evaluation for mobile software based on AADL language and the StrongARM processor. Further, a process for evaluating energy consumption of mobile software is defined based on AADL language. In addition, a tool for evaluating energy consumption is developed. Lastly, a method for evaluating the energy consumption of mobile software is proposed based on AADL language. The experimental results show that the proposed method improves precision compared with existing evaluation method of energy consumption based on AADL.