Fall accidents are critical issues in an aging and aged society. Recently, many researchers developed "pre-impact fall detection systems" using deep learning to support wearable-based fall protection systems for preventing severe injuries. However, most works only employed simple neural network models instead of complex models considering the usability in resource-constrained mobile devices and strict latency requirements. In this work, we propose a novel pre-impact fall detection via CNN-ViT knowledge distillation, namely PreFallKD, to strike a balance between detection performance and computational complexity. The proposed PreFallKD transfers the detection knowledge from the pre-trained teacher model (vision transformer) to the student model (lightweight convolutional neural networks). Additionally, we apply data augmentation techniques to tackle issues of data imbalance. We conduct the experiment on the KFall public dataset and compare PreFallKD with other state-of-the-art models. The experiment results show that PreFallKD could boost the student model during the testing phase and achieves reliable F1-score (92.66%) and lead time (551.3 ms).

In order to improve healthcare services and support clinical professionals, it is important to develop the unobstructive and automatic ADLs monitoring system for healthcare applications. Currently, various works have been developed for the monitoring of daily activities, such as ambulation, kitchen task, food and fluid intake, dressing, and medication intake while only few works paid attention to the housekeeping task. Housekeeping activity is a complex task, generally important for the several clinical assessment tools. In this work, we design and develop a wearable sensor-based activity recognition system recognize housekeeping tasks and classify the activity level. The proposed system achieves 90.67% accuracy for housekeeping tasks recognition, and 94.35% accuracy for activity level classification, respectively. The results of the experiment demonstrate that the system is reliable and fulfills the requirements of the unobstructive, objective, and long-term monitoring system.

Rehabilitation exercise is one of the most important parts in knee osteoarthritis therapy. A good rehabilitation monitoring method provides physiotherapists with performance metrics that are greatly helpful in recovery progress. One of the main difficulties of monitoring and analysis is performing accurate online segmentation of motion sections due to the high degree of freedom (DoF) of human motion. This paper proposes an approach for initial posture classification and online segmentation of rehabilitation exercise data acquired with body-worn inertial sensors. Specifically, we introduce a threshold-based algorithm for initial posture classification and a multi-layer Support Vector Machine (SVM) model for online segmentation. The proposed approach is capable of accurate online segmentation and classification of exercise data. The approach is verified on 10 subjects performing common rehabilitation exercises for knee osteoarthritis, giving initial posture classification accuracy of 97.9% and segmentation accuracy of 90.6% on layer-1 SVM and 92.7% on layer-2 SVM.

According to the WHO report in 2013, the world population aging over 60 years is predicted to increase to 20 million in 2050. Aging comes about many challenges to elders due to their cognitive decline, chronic age-related diseases, as well as limitations in physical activity, vision, and hearing. Recent advances in wearable computing and mobile health technology create new opportunity for ambient assisted living system to help the person perform the activities safely and independently. The activity monitoring of daily living is the core technique of the ambient assisted living system. Several well-known approaches have utilized various sensors for activity recognition such as camera, RFID, infrared detector and inertial sensor. Since the activities are well characterized by the objects, location or hand gesture that are manipulated during their performance on activities of daily living. However, some applications included, e.g. the monitoring of specific tasks and/or movements in a rehabilitation scenario or the classification of dietary intake gestures for an automated nutrition monitoring system, where reliable activity recognition on a more fine-grained level is needed. To fulfill the requirement, we design a hierarchical window approach based on the dynamic time warping algorithm to achieve fine-grained activity recognition, where the template selection and threshold configuration is developed to cope with the ambiguity with similar features. Furthermore, a confidence estimation for the pattern matching is also proposed. The recognition procedure was successfully adapted to the investigated cleaning tasks. The overall performance in precision, recall, and F1-socre is 89.0%, 88.6%, and 88.1% respectively. The results of the experiment demonstrate that the proposed mechanism is reliable and fulfills the requirements of the ambient assisted living.

Patient monitoring systems are gaining their importance as the fast-growing global elderly population increases demands for caretaking. These systems use wireless technologies to transmit vital signs for medical evaluation. In a multihop ZigBee network, the existing systems usually use broadcast or multicast schemes to increase the reliability of signals transmission; however, both the schemes lead to significantly higher network traffic and end-to-end transmission delay. In this paper, we present a reliable transmission protocol based on anycast routing for wireless patient monitoring. Our scheme automatically selects the closest data receiver in an anycast group as a destination to reduce the transmission latency as well as the control overhead. The new protocol also shortens the latency of path recovery by initiating route recovery from the intermediate routers of the original path. On the basis of a reliable transmission scheme, we implement a ZigBee device for fall monitoring, which integrates fall detection, indoor positioning, and ECG monitoring. When the triaxial accelerometer of the device detects a fall, the current position of the patient is transmitted to an emergency center through a ZigBee network. In order to clarify the situation of the fallen patient, 4-s ECG signals are also transmitted. Our transmission scheme ensures the successful transmission of these critical messages. The experimental results show that our scheme is fast and reliable. We also demonstrate that our devices can seamlessly integrate with the next generation technology of wireless wide area network, worldwide interoperability for microwave access, to achieve real-time patient monitoring.

Abstract The proliferation of communication and mobile computing devices and local‐area wireless networks has cultivated a growing interest in location‐aware systems and services. An essential problem in location‐aware computing is the determination of physical locations. RFID technologies are gaining much attention, as they are attractive solutions to indoor localization in many healthcare applications. In this paper, we propose a new indoor localization methodology that aims to deploying RFID technologies in achieving accurate location‐aware undertakings with real‐time computation. The proposed algorithm introduces means to improve the accuracy of the received RF signals. Optimal settings for the parameters in terms of reader and reference tag properties were investigated through simulations and experiments. The experimental results indicate that our indoor localization methodology is promising in applications that require fast installation, low cost and high accuracy. Copyright © 2009 John Wiley & Sons, Ltd.

Nowadays, IP networks are rapidly evolving toward a QoS-enabled infrastructure. The need for packet classification is increasing in accordance with emerging differentiated services. While the new differentiated services could significantly increase the number of rules, it has been demonstrated that performing packet classification on a potentially large number of rules is difficult and has poor worst-case performance. In this work, we present an enhanced tuple pruning search algorithm called "tuple pruning plus" (TPP) for packet classification, which outperforms the existing schemes on the scalability. Our main idea is to simplify the lookup procedure and to avoid unnecessary tuple probing by maintaining the least-cost property of rule through precomputation and the proposed information marker. With extra rules added for information marker, only one tuple access is required in each packet classification. In our experiments, 70 MB DRAM is used to achieve 50 million packets per second (MPPS) for a 1 M-rule set, showing a performance improvement by a factor of 50. We also present a heuristic to further reduce the required storage to about 20 MB. These results demonstrate the effectiveness of the TPP scheme to achieve high speed packet classification

To provide differentiated quality of service, packet classification is important for determining which flow an incoming packet belongs to so as to decide what service quality it should receive. Packet classification is essentially a problem of multidimensional range matching. Tuple space search is a well-known solution based on multiple hash accesses for various filter length combinations. Tuple pruning algorithm is a tuple-based algorithm which is able to achieve good performance in a practical environment; however, its worst-case speed is not guaranteed. We explore the relative property of filters and reorganize the filters through filter conversion. As compared with the tuple pruning algorithm, the proposed scheme can significantly improve the worst-case performance. Experimental results on both real-world and synthetic filter databases show that the worst-case lookup speed of the proposed scheme is 9 to 31 times faster than that of the tuple pruning algorithm.

In next-generation networks, packet classification is important in fulfilling the requirements of multimedia services, including VoIP and VoD. Using pre-defined filters, the incoming packets can be categorized that determines to which forwarding class a packet belongs. Packet classification is essentially a problem of multidimensional range matching. The tuple space search is a well-known solution based on multiple hash accesses for various filter length combinations. The tuple-based algorithm, a rectangle search, is highly scalable with respect to the number of filters; however, it suffers from the memory-explosion problem. Besides, the lookup performance of the rectangle search is not sufficiently fast to accomplish high-speed packet classification. This work proposes an improved scheme to reduce the required storage and realize OC-192 wire-speed forwarding. The scheme consists of two parts. The "Tuple Reduction Algorithm" drastically reduces the number of tuples by duplicating filters. Dynamic programming is used to optimize the tuple reduction and two heuristic approaches are introduced to simplify the optimization process. Furthermore, the "Look-ahead Caching" scheme is presented to improve the lookup performance. The basic idea is to prevent unnecessary tuple probing by filtering out the "un-matched" situation of the incoming packet. The experimental results show that combining the tuple reduction algorithm with look-ahead caching increases the lookup speed by a factor of six while requiring only around one third of the storage. Additionally, an extension of multiple fields to more general filters is addressed.

Hashing is a widely used method to perform fast lookup. Several schemes have been proposed to support Internet lookup that includes IP lookup and packet classification. Rectangular search is a well-known packet classification scheme based on multiple hash accesses for different filter length. It shows good scalability with respect to the number of filters; however, the lookup performance is not satisfactory. For example, through experiments, each packet classification takes about 40 hash accesses in a 100,000-filter database and each hash access may take more than one memory access. Obviously, this is insufficient to provide gigabits throughput. We proposed a novel "lookahead caching" which can significantly improve the performance of the hash-based algorithm. The basic idea is to find out the unmatched case for each incoming packet, thus it is different from the traditional caching mechanism. The experimental results indicate that the proposed scheme can improve the performance by a factor of two. The scheme can be further enhanced using parallel processing.

We present a novel solution to the problem of best matching prefix (BMP) which is required in the IP routing lookup. Our approach is based on the trie-based algorithm. The idea is to prune the trie so that the main-searchable portion of the trie can be fitted into the small on-chip SRAM. The algorithm consists of two parts: level smart-compression trie and trie pruning. In the first part, we present the new trie-based algorithm which can provide flexibility, storage efficiency and incremental update. Moreover, the fixed-size data structure eliminates the complexity of the memory management. Secondly, we define the "core" and present how to apply the concept "trie pruning" to achieve fast IP routing lookup. With the currently popular platform, the proposed scheme can provide 40 MPPS without any assumption. While considering route flaps, the performance degrades by only 0.01% with 4000 BGP updates per 30 seconds.

Nowadays, the commonly used table lookup scheme for IP routing is based on the so-called classless interdomain routing (CIDR). With CIDR, routers must find out the best matching prefix (BMP) for IP packet forwarding, which complicates the IP lookup. Since the IP lookup performance is a major design issue for the new generation routers, we investigate the properties of the routing table and present a new IP lookup scheme. By adopting the extreme compression, the size of the forwarding table can be compressed to 360 Kbytes for a large routing table with 58,000 routing entries. It can accomplish one IPv4 route lookup within maximum nine memory accesses. Consequently, the data structure for the incremental update is introduced. With the new data structure, the proposed scheme can accomplish one route update within 500 ns. Even under the circumstance with 4000 route updates per second, the performance degradation is as low as 0.2 %.

The construction of routing tables has been studied extensively. Although existing work has certain advantages, it either uses complicated data structures which result in large storage requirements and high complexity for updating/building the forwarding table, or it is not scalable to fit in IPv6. Lampson et al. (1999) proposed an IP lookup algorithm which performs binary search on prefixes (BSP). The algorithm is attractive, even for IPv6, because of its bounded worst-case memory requirement. For achieving fast forwarding, the cost is the slowing down of insertion. Although this can be justified, the performance of routing-table reconstruction in BGP is too time-consuming to handle frequent route updates. We propose a fast forwarding table construction algorithm, which can handle more than 4,000 route updates per second. Moreover, it is simple enough to fulfil the need of fast packet forwarding. By using the modified multiway search tree, we can further reduce the depth of the tree and eliminate storage for pointers. This reduces the forwarding table size and shortens the lookup time.

A major design issue for the next generation routers is the IP lookup mechanism. The router needs to perform a longest prefix matching on the address lookup for each incoming packet to determine the next hop. Currently, the process is done in software and has become a major performance bottleneck of the router. We propose a fast IP lookup mechanism in which the forwarding table is small enough to fit in an SRAM with very low cost. It also can be implemented in hardware using the pipeline technique. A large routing table with 45,000 routing prefixes can be compressed to a forwarding table with about 430 kbytes in sine by using our proposed method. In the worst case, the number of memory accesses for a lookup is three. When implemented in pipeline technique, the proposed mechanism can achieve one routing lookup per memory access. With current 10 ns SRAM, this mechanism provides approximately 100 million routing lookups per second. Furthermore, the lookup speed can be improved linearly through the speedup of the memory access.