Vietnamese, a low-resource language, is typically categorized into three primary dialect groups that belong to Northern, Central, and Southern Vietnam.However, each province within these regions exhibits its own distinct pronunciation variations.Despite the existence of various speech recognition datasets, none of them has provided a fine-grained classification of the 63 dialects specific to individual provinces of Vietnam.To address this gap, we introduce Vietnamese Multi-Dialect (ViMD) dataset, a novel comprehensive dataset capturing the rich diversity of 63 provincial dialects spoken across Vietnam.Our dataset comprises 102.56 hours of audio, consisting of approximately 19,000 utterances, and the associated transcripts contain over 1.2 million words.To provide benchmarks and simultaneously demonstrate the challenges of our dataset, we fine-tune state-of-the-art pre-trained models for two downstream tasks: (1) Dialect identification and (2) Speech recognition.The empirical results suggest two implications including the influence of geographical factors on dialects, and the constraints of current approaches in speech recognition tasks involving multidialect speech data.Our dataset is available for research purposes 1 .

Eye blink is a quick action of closing and opening of the eyelids. Eye blink detection has a wide range of applications in human computer interaction and human vision health care research. Existing approaches to eye blink detection often cannot suit well resource-limited eye blink detection platforms like Smart Glasses, which have limited energy supply and typically cannot afford strong imaging and computational capabilities. In this paper, we present an efficient and robust eye blink detection method for Smart Glasses. Our method first employs an eigen-eye approach to detect closing-eye in individual video frames. Our method then learns eye blink patterns based on the closing-eye detection results and detects eye blinks using a Gradient Boosting method. Our method further uses a non-maximum suppression algorithm to remove repeated detection of the same eye-blink action among consecutive video frames. Experiments with our prototyped smart glasses equipped with a low-power camera and an embedded processor show an accurate detection result (with more than 96% accuracy) on video frames of a small size of 16 × 12 at 96 fps, which enables a number of applications in health care, driving safety, and human computer interaction.

In this paper, we propose a framework to support the bridging of applications and computer-vision based sensor networks. We argue that the semantic gap, the difference between the data collected in a sensor network and the information needed by the application, in video-based sensor networks can only be addressed by providing systems support in such a way that allows users and computing systems to meet in the middle. We first outline the vision of the system that we are working towards. We then describe initial experiments that we have conducted using a functional component of the system applied to real-world video data that is being collected by intelligent transportation systems researchers.

As sensor networking technologies continue to develop, the notion of adding large-scale mobility into sensor networks is becoming feasible by crowd-sourcing data collection to personal mobile devices. However, tasking such networks at fine granularity becomes problematic because the sensors are heterogeneous and owned by users instead of network operators. In this paper, we present Zoom, a multi-resolution tasking framework for crowdsourced geo-spatial sensor networks. Zoom allows users to define arbitrary sensor groupings over heterogeneous, unstructured and mobile networks and assign different sensing tasks to each group. The key idea is the separation of the task information ( what task a particular sensor should perform ) from the task implementation ( code ). Zoom consists of (i) a map, an overlay on top of a geographic region, to represent both the sensor groups and the task information, and (ii) adaptive encoding of the map at multiple resolutions and region-of-interest cropping for resource-constrained devices, allowing sensors to zoom in quickly to a specific region to determine their task. Simulation of a realistic traffic application over an area of 1 sq. km with a task map of size 1.5 KB shows that more than 90% of nodes are tasked correctly. Zoom also outperforms Logical Neighborhoods, the state-of-the-art tasking protocol in task information size for similar tasks. Its encoded map size is always less than 50% of Logical Neighborhood's predicate size.

Current sensor networks have limited use after deployment due to inability to dynamically update sensing tasks in the networks. Assigning multiple tasks to multiple sensor groups in a sensor network deployed over a large geographic region at fine granularity is problematic because the sensors are heterogeneous, potentially mobile, and owned by users instead of network operators. We demonstrate Zoom [1], a multi-resolution tasking framework for crowdsourced geo-spatial sensor networks. The key idea in Zoom is decoupling the task specification from the task implementation using a spatial 2-dimensional representation of the tasking region (e.g., maps). Zoom allows users to define sensor groups over heterogeneous, unstructured and mobile networks and assign different sensing tasks to each group. To the best of our knowledge, this is the first work to propose a map based approach for mobile sensing systems.

The ability to create super high-resolution video is becoming relative easy to do either through a single high-definition video camera or panoramic video that automatically stitches multiple views together. As an example of the former, the motion picture industry now has 6000 × 4000 pixel full-rate video cameras available. This means that supporting region-of-interest cropping will become more important in the future. In this article, we propose a mechanism to support region-of-interest adaptation of stored video. The proposed approach creates a compression-compliant stream (e.g., MPEG-2), while still allowing it to be cropped. Fortunately, video standards like MPEG-2 specify the format of a compliant stream, and not the algorithm to get there. As a result, there is an opportunity to allow system researchers and implementers ways to optimize for applications. We show various fundamental tradeoffs that are made in order to support region-of-interest cropping with super high-resolution video which we received from a local motion-picture firm.

Current feature tracking frameworks in sensor networks exploit advantages of either mobility, where mobile sensors can provide micro scale information of a small sensing area or numerical models that can provide macro scale information about the environment but not both. With the continual development of sensor networks, mobility becomes an important feature to integrate next generation sensing systems. In addition, recent advances in environmental modeling also allow us to better understand basic behavior of the environment. In order to further improve existing sensing systems, we need a new framework that can take advantages of existing fixed sensor networks, mobile sensors and numerical models. We develop CoTrack, a Collaborative Tracking framework, that allows mobile sensors to cooperate with fixed sensors and numerical models to accurately track dynamic features in an environment. The key innovation in CoTrack is the incorporation of numerical models at different scales and sensor measurements to guide mobile sensors for tracking. The framework includes three components: a macro model for large-scale estimation, a micro model for locale estimation of specific features based on sensor measurements, and an adaptive sampling scheme that guides mobile sensors to accurately track dynamic features. We apply our framework to track salinity intrusion in the Columbia River estuary in Oregon, United States. Our framework is fast and can reduce tracking error by more than half compared to existing data assimilation and state-of-the-art numerical models.

The ability to create very high-resolution video is becoming relative easy to do today, either through a single high definition video camera or panoramic video stitched from multiple cameras. This means that supporting region-of-interest cropping will become more important in the future. In this paper, we propose a mechanism to support region-of-interest adaptation of stored video. The proposed approach creates a compression compliant stream (e.g., MPEG-2), while still allowing it to be cropped. Using data from a motion-picture industry video camera, we show that the proposed approach, while adding a small amount of overhead, can be used to efficiently crop large resolution video into a smaller stream.