Latency-sensitive edge video analytics applications require rapid responses for real-time decision-making, driving the demand for efficient object detection pipelines. Conventional pipelines transmit and process full frames, overlooking redundancy in videos and leading to unnecessary resource consumption. Existing block-wise conditional execution methods mitigate this issue by processing only informative blocks. However, they treat all informative blocks equally and fail to further categorize these blocks. To address this limitation, we propose BlockHybrid, an edge video analytics framework designed to accelerate object detection pipelines by hybrid block-wise execution. Specifically, BlockHybrid classifies blocks into hard or easy blocks using a policy network. Hard blocks are transmitted and processed by a block-wise detector on the server, while easy blocks are handled by an efficient tracker locally on the camera, reducing redundant computation and communication. Extensive experiments demonstrate that BlockHybrid can achieve 8.8%–31.5% higher local execution speed and comparable detection accuracy compared to state-of-the-art methods, and accelerate end-to-end processing—including camera-to-server communication—by 31.5%–39.1% in a real-world testbed.

OBJECTIVES: Effective communication is crucial in healthcare, and for patients with a non-English language preference (NELP), professional interpreters are recognized as the gold standard in supporting bidirectional communication. However, interpreters are not always readily available, prompting the exploration of other options for translation and interpretation. The recent developments in artificial intelligence-based neural network translation tools, namely neural machine translation (NMT) may enable robust interpretation and translation. MATERIALS AND METHODS: We conducted a systematic review (SR) to evaluate the literature on NMT for this purpose. We did a comprehensive search of several databases with guidance from a professional librarian. The search was limited to the year 2000 onwards and English language. Title and abstract screening and full-text review were independently conducted by two reviewers with conflicts resolved by a third reviewer. RESULTS: 2867 studies were identified with 10 studies included in the final analysis. Among these, six evaluated interpretation in real or simulated clinical settings and four examined translation of discharge materials. Google Translate and ChatGPT were assessed in several studies. Accuracy differed by language, with low-resource languages performing worse. DISCUSSION: NMT technologies in healthcare have several advantages including broad language accessibility and potential cost savings for institutions. Despite improved accuracy of these novel tools, due to possible critical errors NMT tools are not yet ready for widespread clinical use. CONCLUSION: Future studies should focus on optimizing evaluation methods as well as how best to integrate these technologies into real-time clinical settings.

Multi-object tracking (MOT) is the “killer app” of edge video analytics. Deploying MOT pipelines for live video analytics poses a significant system challenge due to their computation-intensive nature. In this paper, we propose FastTuner, a model-agnostic framework that aims to accelerate MOT pipelines by adapting frame resolutions and backbone models. Unlike prior works that utilize a separate and time-consuming online profiling procedure to identify the optimal configuration, FastTuner incorporates multi-task learning to perform configuration selection and object tracking through a shared model. Multi-resolution training is employed to further improve the tracking accuracy across different resolutions. Furthermore, two workload placement schemes are designed for the practical deployment of FastTuner in edge video analytics systems. Extensive experiments demonstrate that FastTuner can achieve 1.1%–9.2% higher tracking accuracy and 2.5%–25.5% higher speed compared to the state-of-the-art methods, and accelerate end-to-end processing by 1.7%–22.5% in a real-world testbed consisting of an embedded device and an edge server.

Tracking people in multi-camera surveillance systems is challenging due to disparate perspectives, large volumes of data, and high computation demands. This paper presents a distributed cooperative pipeline for pedestrian tracking that exploits the spatial and temporal redundancy within and across the video feeds from multiple synchronized cameras. It consists of three key components: 1) a lightweight policy network trained online in a self-supervised manner on each camera, 2) a sparse backbone processing unit purpose-built for parallel processing of selected regions of all cameras, and 3 an online clustering algorithm for object association. Utilizing online distributed reinforcement learning, the fully end-to-end trainable pipeline can accelerate any tracking-by-detection method by reducing detection costs across multiple perspectives. MVSparse has been evaluated using two multi-camera multi-target pedestrian tracking datasets, WildTrack and MultiviewX. It reduces the amount of processed regions by up to 52% and 39% with only moderate degradation of 1% and 0.1% in tracking accuracy on the two datasets, respectively. On a real-world testbed comprising four NVIDIA Jetson TX2 and a GPU server, MVSparse accelerates the end-to-end process and reduces the communication overheads by 1.88 and $1.60 X$ with only 2.27% and 3.17% degradation in tracking accuracy on the two datasets, respectively

Multi-object tracking (MOT) is a vital component of intelligent video analytics applications such as surveillance and autonomous driving. The time and storage complexity required to execute deep learning models for visual object tracking hinder their adoption on embedded devices with limited computing power. In this paper, we aim to accelerate MOT by transferring the knowledge from high-level features of a complex network (teacher) to a lightweight network (student) at both training and inference times. The proposed AttTrack framework has three key components: 1) cross-model feature learning to align intermediate representations from the teacher and student models, 2) interleaving the execution of the two models at inference time, and 3) incorporating the updated predictions from the teacher model as prior knowledge to assist the student model. Experiments on pedestrian tracking tasks are conducted on the MOT17 and MOT15 datasets using two different object detection backbones YOLOv5 and DLA34 show that AttTrack can significantly improve student model tracking performance while sacrificing only minor degradation of tracking speed.