INTRODUCTION:Video-based automatic motion analysis has been employed to identify infant motor development delays. To overcome the limitations of lab-recorded images and training datasets, this study aimed to develop an artificial intelligence (AI) model using videos taken by mobile phone to assess infants' motor skills.METHODS:A total of 270 videos of 41 high-risk infants were taken by parents using a mobile device. Based on the Pull to Sit (PTS) levels from the Hammersmith Motor Evaluation, we set motor skills assessments. The videos included 84 level 0, 106 level 1, and 80 level 3 recordings. We used whole-body pose estimation and three-dimensional transformation with a fuzzy-based approach to develop an AI model. The model was trained with two types of vectors: whole-body skeleton and key points with domain knowledge.RESULTS:The average accuracies of the whole-body skeleton and key point models for level 0 were 77.667% and 88.062%, respectively. The Area Under the ROC curve (AUC) of the whole-body skeleton and key point models for level 3 were 96.049% and 94.333% respectively.CONCLUSIONS:An AI model with minimal environmental restrictions can provide a family-centered developmental delay screen and enable the remote monitoring of infants requiring intervention.

Monocular 2D multi-person pose estimation in videos is essential for applications such as surveillance, action recognition, kinematics analysis, and medical diagnosis. Ex-isting state-of-the-art offsets-based methods extract temporal features from offsets in con-secutive predicted rough skeletons for better preciseness in fine-tuned the skeletons. How-ever, the precision of existing single image-based models of rough skeleton prediction, such as HRNet, dropped due to shifting of target persons in propagated bounding boxes and resulted in inconsistent estimated poses in consecutive frames. To conquer this prob-lem, we proposed an Ordinal Depth-Guided-Convolutional Neural Network (ODG-CNN) to address the issue. The proposed ordinal depth guides the Ordinal Depth-Guided Block (ODGB) in the ODG-CNN to reweight features for target persons in bounding boxes. Ex-periment results on the PoseTrack 2018 dataset indicate that the proposed ODG-CNN achieves the highest performance in terms of mean Average Precision (mAP). The pro-posed ODG-CNN is suited for applications, such as use of telehealth for early detection and intervention of developmental delays in children, which needs high accuracy of video -based estimated poses.

Real-time control solutions for unmanned aerial vehicles (UAVs) have attracted great interest in recent years. Most existing control methods use Wi-Fi technology. While Wi-Fi is inexpensive and easy to use, it has only a limited transmission range. Thus, 4G/5G cellular networks have been proposed as an alternative enabling technology. This study focuses on the problem of improving the appropriateness of the control commands sent by the ground control station (GCS) to the UAV over the control and nonpayload communication (CNPC) link of the UAV through the cellular network. To satisfy the low-latency requirement of the CNPC link, multiaccess edge computing (MEC) technology is leveraged to collocate the GCS and base station. The effectiveness of the proposed edge-based approach is demonstrated by conducting experiments on two LTE platforms with different MEC deployment methods. An edge-enabled UAV control solution is proposed in which each end-to-end control delay in the UAV-GCS system is estimated based on the preceding delay such that the location of the UAV at the moment it receives the control command from the GCS can be predicted in advance and taken into consideration by the GCS when formulating an appropriate control decision. To this end, an analytical modeling method is proposed for estimating the expected error range of each control delay based on a bimodal distribution approximation of the empirical control delays observed at the UAV. Finally, an event-driven simulator is developed to confirm the accuracy of the analytical predictions of the control delay based on the expected error between consecutive delays.

Atrial Fibrillation (AF) is the most common arrhythmia type that affects patients today. Detecting and classifying a patient's electrocardiogram (ECG) beats, especially the supra -ventricular ectopic beats (SVEB) class, can help assess if the patient has high possibilities of AF/atrial flutter in the future. Detecting the SVEB class considered more difficult than the other classes. Related works show low classification (prediction) performance, in terms of sensitivity, F1 score, and G score, for detecting the SVEB class in a single-lead ECG. This work focuses on designing an arrhythmia beats detection method using single-lead ECG data with a patient-specific training model design, and does classification based on the AAMI standards. This work aims at achieving high classification performance in the SVEB class and still meets the real-time ECG classification requirement. The proposed method uses Empirical Mode Decomposition (EMD) with resampling (EMDR), which resamples only the first Intrinsic Mode Function (i.e., IMF 1) as a main input, for the pro-posed EMDR-LSTM (Long Short-Term Memory) architecture. In contrast to the related works that use two separate models with one or two LSTM layers for each input, we de-signed a novel LSTM model architecture that only uses a single model with one LSTM layer for each input. The proposed LSTM architecture is suited for our preprocessing method, EMDR, and can enhance the SVEB classification performance. To the best of our knowledge, the proposed EMDR-LSTM is the first one that uses resamples first IMF in LSTM that classifies arrhythmia using single-lead ECG data based on the AAMI standards. Compared to representative related works, experiment results show that the proposed EMDR-LSTM achieves the highest classification performance, in terms of the following performance metrics: accuracy, sensitivity, positive predictivity, and F1 and G scores, for the SVEB class in all datasets used. In addition, although the proposed EMDR-LSTM has higher preprocessing cost and higher computational complexity in terms of MACs (multi-ply-accumulate operations), it has lower standard deviations of the performance metrics and lower inference time, which are important performance metrics for real time or time -critical applications, e.g., ECG medical monitoring applications, compared to the repre-sentative related works.

Unmanned aerial vehicle (UAV) has been an emerging technology used for various applications. Enabling wireless base station (e.g., Wi-Fi AP) on UAV can be one of promising UAV applications. In this work, we focus on a performance fairness issue on ground Wi-Fi users and aim to maximize the minimum throughput performance among them. To this end, we propose a solution, designated UAV-FAP (UAV with user Fairness-driven AP), to drive UAV-AP to reach a location that can offer max-min throughput performance. It employs a heuristic-based hierarchical search method to search for the target location efficiently. By conducting experiments on a real UAV-AP platform, we first study the issues of antenna pointing direction and vertical/horizontal UAV-AP placement, and then collect Wi-Fi throughput statistics for a trace-driven simulator in the evaluation. The evaluation result shows that UAV-FAP can outperform the default case, where the UAV-AP stays at the center of a serving area, by 2.0%-21.5% throughput gains; in some cases, it takes as small as only 7% of the moving distance needed by a fine-grained exhaustive search to reach target max-min locations.

Unmanned aerial vehicles (UAVs) are being investigated to substitute for labor in many indoor applications, e.g., asset tracking and surveillance, where the global positioning system (GPS) is not available. Also, emerging autonomous UAVs are expected to land in indoor parking aprons automatically. Such GPS-denied environments require alternative non-GPS positioning methods. Although there have been some vision-based solutions for UAVs, they perform poorly in the scenes with bad illumination conditions or estimate only relative locations but not global positions. Other common indoor localization methods do not cover UAV factors, such as low power and flying behaviors. To this end, we propose a practical non-GPS positioning system for UAVs, named WBF-PS (WiGig Beam Fingerprinting based Positioning System), using low-power, off-the-shelf WiGig devices. We formulate a 3-dimensional beam fingerprint for the positioning by leveraging the diversity of available transmitter/receiver beams and the link quality. To augment the positioning accuracy, we not only use a weighted k-nearest neighbors algorithm to overcome partial fingerprint inaccuracy but also apply the particle filtering technique into considering the UAV motion. We prototype and evaluate WBF-PS on a UAV platform. The result shows that the positioning errors at the 90th percentile are below 1 m in various cases.

With the use of the Internet, mobile platforms, online commerce, and social media services, the footprints of human behavior can be easily recorded in the digital world, which generates data on an extremely large scale. Twitter as a big data social network becomes one of the most important sources for capturing up-to-date events happened in the world. Deriving topics from Twitter is important for various applications, such as situation awareness, market analysis, content filtering, and recommendations. However, topic derivation with high purity in Twitter is hard to achieve because tweets are limited to 140 characters. Previous works on topic derivation in Twitter suffer from low purity. In this paper, we propose corpus-based topic derivation (CTD) approach that combines a Twitter corpus and LF-LDA, which is a text processing model to identify topics and clusters of similar hashtags. We use asymmetric topic LF-LDA to obtain better purity of topics. Compared to intJNMF, a representative related work, the purity (F-measure) of our proposed CTD increases from 5.26% (27.81%) to 11.32% (34.28%) for 20 to 100 topics. We also propose a timestamp-based popular hashtags prediction (TPHP) approach by creating trending hashtags lists (THLs), which are lists of hashtags used by many users and make use of timestamps in tweets. We use the edit distance to find the difference between consecutive THLs. Then the difference can be used to calculate volatilety to find how people react to real world events. Compared to Hybrid+, a representative related work, the mean average precision (MAP) of our TPHP increases by 19.45% (week-day), 15.08% (week-week) and 16.95% (month-week).

The multi-controller architecture is a must to improve scalability and reliability of software defined networks (SDNs). A hierarchical multi-controller architecture allows a global controller to handle rare events, such as routing events from local controllers. A failover mechanism, which includes failure detection and recovery, is necessary for the multi-controller architecture in case of a controller failure. Existing controller failure detection and recovery mechanisms are performed in local controllers, which may degrade the performance of the local controllers. In controller failure recovery, existing mechanisms did not consider the time-varying load standard deviation metric, which may be used to reduce the flow setup time. To address these problems, we propose a Simulated Annealing-based Efficient Failover (SAEF) mechanism, run as an app in global controllers, to relieve the burden of local controllers for high availability hierarchical SDN controllers. The SAEF uses multiple global controllers to detect a local controller failure and compute a switch migration plan to redo switch-controller association for controller failure recovery. In controller failure detection, global controllers exchange their failure detection results of local controllers to jointly make a local controller failure decision to reduce the mistake rate. In controller failure recovery, the SAEF considers two metrics: switch-controller propagation delay and load standard deviation among local controllers with adaptive weights to adapt to time-varying local controllers' loads. In addition, in our design, a switch is connected to two local controllers so that asynchronous messages, such as packet-in and port-status messages, will not be lost during controller failover. Experiment results using EstiNet show that, the SAEF can achieve efficient failover with the best load balance result in term of load balance metric (LBM) of 1.035, while the best LBM of related works is 1.087. The SAEF also has the lowest flow setup time of 7.837 ms, while the best of related works is 10.418 ms. The lower flow setup time implies that the SDN data plane can receive the response message of a packet-in message faster.

Most existing datacenters still use Equal-Cost Multi-Path (ECMP) to achieve network load balancing. But such a stateless load balancing mechanism is not able to react adaptively when network congestion occurs. It will cause datacenters unable to make good use of network resources under the unbalanced network state. The main reason most datacenters still using ECMP is that the cost of realizing the state-of-art load balancing technologies is too high since all switches in datacenters need to be replaced with programmable switches, such as P4 switches. In this paper, we propose a cost-effective congestion-aware load balancing (CCLB) scheme that only needs to replace a portion of switches with programmable switches to achieve congestion-aware load balancing. CCLB makes good use of Explicit Congestion Notification (ECN) in the IP layer to be aware of network congestion and uses flowlet switching that slices large flows into small sub-flows to achieve load balancing. Experiment results show that the average flow completion time of the proposed CCLB is 27% shorter than that of ECMP in large flows. Compared to a classical congestion-aware mechanism, HULA, the average flow completion time of the proposed CCLB is slightly longer, but our switch replacement cost is much lower than HULA's.

Commercial OpenFlow-enabled legacy switches implement flow tables with Ternary Content Addressable Memory (TCAM) to perform high speed flow matching process. However, the size of TCAM-based flow table is restricted to few thousands of entries due to manufacturing cost and high power consumption. In this paper, we propose the Flow Entry Agent which can integrate heterogeneous flow tables, such as ASIC-side MAC/IP/ACL tables and CPU-side SRAM-based tables to increase the storage capacity without modifying ASIC. Besides, the Agent will compress the CPU-side tables and perform flow matching process with these tables. Simulation results show that our Agent can not only increase the capacity of flow table to 11.38x to 65.24x but also reduce 80.99% to 87.36% of Packet_In messages with additional 1 MB CPU-side SRAM.

In the software defined networks (SDN), the flow table of an OpenFlow switch is usually implemented by the ternary content addressable memory (TCAM). The TCAM has limited capacity which may cause the flow table overflow problem. To address the problem, related works perform a rule replacement scheme according to either the estimated next-packet arrival time of each flow or the hit count of each rule; however, it may remove a rule whose packets will soon arrive at a switch, especially when the network is unstable. This paper proposes a novel TCAM management scheme named in-switch rule caching and replacement (IRCR) to better deal with the flow table overflow problem. The IRCR replaces a rule according to the expected number of incoming matched flows (EIMF), which can be derived from the inter-arrival time distribution models of the corresponding flows. The uniqueness of the IRCR is that it is designed as an in-switch application to avoid link delays between controllers and switches. Using either captured real network traffic or uniform distribution traffic, the IRCR outperforms related works for both exactly matched rules and wildcard rules in terms of hit ratios. In addition, the IRCR has little overhead to the switch.

In this paper, we present a topology-aware network coding (TANC) technique for wireless multicast systems aiming at reducing packet retransmissions in a lossy propagation environment. In addition to radio channel impacts, packet loss may result from traffic congestion, thereby yielding different retransmission requests. In such a wireless network coding multicast environment, what is the best packet combination to minimize the packet retransmissions for a number of users? We make progress on this problem by taking advantage of the network topology information to select the appropriate lost encoded packets in the encoding phase. We leverage the capability of software-defined networking controller to construct the global network topology in the proposed TANC scheme. Experimental results show that the proposed TANC scheme can reduce 52.4% of packet retransmissions compared to the existing random linear network coding in different environments, including various numbers of uncoded packets, network nodes, and packet loss probabilities.

Software-defined networking (SDN) apps are developed to support various functions (e.g., traffic engineering, routing, security, etc.) for SDN networks. Their operations rely on the APIs offered by the control plane. They may be compromised or designed to be malicious by third parties. Though there have been many studies against malicious apps, they only restrict the APIs used by them with coarse-grained controls. In this work, we seek to show that some malicious flow entries cannot be detected or prevented by current defenses. They may impede the operations of control-plane services or hinder packets from being forwarded correctly in the data plane. To show their negative impact, we devise two attacks, topology spoofing and forwarding-based DoS, as well as examine their damage and analyze root causes. We then propose a context-aware, event-based anomaly detection (CEAD) framework to defend against the malicious flow entries. It provides more fine-grained controls over the flow entries set by apps. Different from other studies, it does anomaly detection by examining the context correlation between an event, the app registering it, and the flow entries set by the app for the event. Our evaluation results show that the CEAD can detect all the malicious flow entries in our given cases, and confirm its scalability with negligible overhead at increasing TCP connection attempt rates.

With numerous emerging internet of things ( IoT) devices, they generate big data. The big data transmitted to the cloud or fog will consume massive network bandwidth. This may result in the IoT network easily encountering network congestion. Moreover, there are IoT applications that need to transfer multimedia data with multiple quality of service (QoS) requirements. A state-of-the-art, MINA, intends to meet multiple QoS requirements of IoT applications; however, it is unable to guarantee QoS requirements of high-priority IoT applications and it is also unable to adapt to the current network status. To conquer the above problems, we propose an application-aware QoS routing algorithm (AQRA) for SDN-based IoT networking to guarantee multiple QoS requirements of high-priority IoT applications and to adapt to the current network status for better routing paths. Evaluation results have shown that, the AQRA has better fitness ratios of QoS requirements compared to MINA while multiple QoS requirements of high-priority IoT applications are guaranteed. The AQRA improves the average end-to-end flow performance by 10.75%, 11.88% and 10.82% compared to MINA in terms of delay, jitter and packet loss rate, respectively. The AQRA improves the standard deviation of end-to-end flow performance by 14.37%, 17.95% and 14.28% compared to MINA in terms of delay, jitter and packet loss rate, respectively. In addition, the runtime of the AQRA is 38.56% shorter than that of MINA.

Load balancing among multiple controllers is a critical issue in the software-defined networking (SDN), since traffic dynamics prevent control-plane loads from being evenly distributed among controllers. Load imbalance may cause some controllers to be overloaded while the other controllers are still underutilized. Though there have been several proposed solutions, theydo not consider the convergence time of load balancing. We thus propose a fast-converging loadbalancing (FCLB) mechanism that seeks to achieve fast convergence for balancing loads (i.e., assigning switches) among the SDN controllers. Fast convergence can rapidly release controllersfrom being overloaded, thereby shortly recovering network performance and preventing some unanticipated results (e.g., controller crash). Searching for the optimal solution from a large set of switch-controller combinations can lead to a large delay, which can aggravate negative impacts from overloading. We thus leverage the genetic algorithm to lind a near-optimal solutionOur simulation results show that FCLB has at least 20.7% faster convergence time than the other mechanisms while achieving better load balancing performance.

Commercial OpenFlow-enabled legacy switches implement flow tables with Ternary Content Addressable Memory (TCAM) to perform high-speed flow matching process. However, the size of TCAM-based flow table is restricted to few thousands of entries due to manufacturing cost and high power consumption. In our previous research, we proposed Flow Entry Agent to integrate Off-Chip tables to increase the storage capacity without modifying Application Specific Integrated Circuit (ASIC). In this paper, we propose an efficient matching mechanism for Off-Chip tables. Simulation results show that our mechanism significantly increase the efficiency of flow matching process.

The limited size of the flow table, usually implemented by ternary content addressable memory (TCAM), may result in the flow table overflow problem, which becomes severer in software defined networks (SDN). To conquer this problem, this paper proposes an in-switch dynamic flow aggregation (IDFA) mechanism, which can be dynamically triggered. In the IDFA, redundant flow entries are inserted to speed up flow aggregation convergence time. The uniqueness of the proposed IDFA is that it resides in an OpenFlow switch, instead of acting as an application in an SDN controller, which avoids link delays between the controller and switches. This paper also proposes two novel techniques, degradation and repermutation, to aggregate flows effectively while keeping semantic equivalence. Evaluation results show that, the average compression ratio of the IDFA in a fat-tree topology under a LAN (random IP) environment is 26.3% (43%) better than that of the Flow Table Reduction Scheme (FTRS), a representative related work. The average flow aggregation convergence time of the IDFA in the fat-tree topology is 84% shorter than that of the FTRS. In addition, the occurrence probability of flow table overflow in the fat-tree topology is 26.71 % lower than that of the FTRS.

Software-Defined Networking (SDN) benefits from the development flexibility of control-plane applications (SDN-Apps), which allows third parties to make contributions. Such flexibility may expose SDN networks to security threats, since SDN-Apps may be malicious or prone to implementation bugs. These buggy/malicious SDN-Apps may contaminate the data plane with abnormal network actions, which may not be prevented before they are committed to the data plane. This contamination may lead to network crash or poor network performance. We thus present ReSDN, a lightweight solution for data-plane state recovery, to recover an SDN data plane from a contaminated state. It requires neither switch modification nor the intervention of SDN-Apps, both of which current recovery solutions rely on. It leverages the concept of FP-tree (Frequent Pattern tree) to maintain the dependency of event transactions and network actions to achieve correct recovery. Our evaluations validate the viability of our ReSDN design, and show that it can recover more than twice as fast as the other type of recovery approach, rollback recovery.

WiGig enables wireless multi-gigabit communication over 60GHz band. However, its usage scenarios may be constrained by two major limitations: line-of-sight propagation and very short transmission range. We seek to boost the WiGig's usability by using WiFi to complement its limitations in triband (2.4/5/60GHz) wireless systems. Our goal is to let a triband client have multimedia services at the WiGig's very high speed without any hassle. When the WiGig link is down or performs bad, not only can the client temporarily handover to WiFi without service interruption, but its ongoing multimedia services can also adapt to the WiFi's slower link. Though the IEEE 802.11ad standard has proposed an FST (Fast Session Transfer) mechanism to support handover operations at the link layer, it does not satisfy our goal due to two reasons. First, it does not specify when to performWiGig/WiFi handovers. Second, it is not application-aware to achieve the service adaptation. To this end, we design and implement an application-aware, seamless WiGig/WiFi handover solution above the network layer. It ensures timely handover trigger for the WiGig's abrupt link interruption, keeps service continuity during handovers, and adapts multimedia service qualities to differentWiGig/WiFi links. Our demo confirms its viability. We show that a video streaming service at the client is not interrupted during WiGig/WiFi handovers, which are triggered by mobility or the WiGig's signal blockage, but smoothly switches between different resolutions according to different links.

Changing the whole network into a software defined network (SDN) is impractical due to high replacement cost. Therefore, there exist SDN hybrid networks, where SDN switches and legacy switches coexist at the same time. In this paper, we propose an SDN hybrid network architecture which can discover existence of legacy switches by using the Spanning Tree Protocol and thus have a global view of the SDN hybrid network. We also enable OpenFlow switches to cooperate with legacy switches by using the Learning Bridge Protocol without requiring any modification on legacy switches. By utilizing the characteristics of SDN, SDN applications can dynamically find routing paths according to pre-defined QoS requirements and current network status. We also propose a simulated annealing based QoS-aware routing (SAQR) algorithm which can adaptively adjust weights of delay, loss rate and bandwidth requirements in a cost function to find the best fit path according to QoS requirements. We evaluate the proposed SAQR in a simulated SDN hybrid network which runs applications with different QoS requirements. Simulation results show that the SAQR performs better than related work MINA in terms of the fitness ratios of delay, loss rate and bandwidth, with 88%, 90.8% and 86.5% of flows meeting their respective QoS requirements, in contrast to MINA, with only 63%, 82.4% and 87.5% of flows meeting their respective QoS requirements.