Learning Online Policies for Person Tracking in Multi-View Environments

In this paper, we introduce MVSparse, a novel and efficient framework for cooperative multi-person tracking across multiple synchronized cameras. The MVSparse system is comprised of a carefully orchestrated pipeline, combining edge server-based models with distributed lightweight Reinforcement Learning (RL) agents operating on individual cameras. These RL agents intelligently select informative blocks within each frame based on historical camera data and detection outcomes from neighboring cameras, significantly reducing computational load and communication overhead. The edge server aggregates multiple camera views to perform detection tasks and provides feedback to the individual agents. By projecting inputs from various perspectives onto a common ground plane and applying deep detection models, MVSparse optimally leverages temporal and spatial redundancy in multi-view videos. Notably, our contributions include an empirical analysis of multi-camera pedestrian tracking datasets, the development of a multi-camera, multi-person detection pipeline, and the implementation of MVSparse, yielding impressive results on both open datasets and real-world scenarios. Experimentally, MVSparse accelerates overall inference time by 1.88X and 1.60X compared to a baseline approach while only marginally compromising tracking accuracy by 2.27% and 3.17%, respectively, showcasing its promising potential for efficient multi-camera tracking applications.

MirrorCalib: Utilizing Human Pose Information for Mirror-based Virtual Camera Calibration

In this paper, we present the novel task of estimating the extrinsic parameters of a virtual camera with respect to a real camera with one single fixed planar mirror. This task poses a significant challenge in cases where objects captured lack overlapping views from both real and mirrored cameras. To address this issue, prior knowledge of a human body and 2D joint locations are utilized to estimate the camera extrinsic parameters when a person is in front of a mirror. We devise a modified eight-point algorithm to obtain an initial estimation from 2D joint locations. The 2D joint locations are then refined subject to human body constraints. Finally, a RANSAC algorithm is employed to remove outliers by comparing their epipolar distances to a predetermined threshold. MirrorCalib is evaluated on both synthetic and real datasets and achieves a rotation error of 0.62{\deg}/1.82{\deg} and a translation error of 37.33/69.51 mm on the synthetic/real dataset, which outperforms the state-of-art method.

Efficient Rotating Synthetic Aperture Radar Imaging via Robust Sparse Array Synthesis

Rotating Synthetic Aperture Radar (ROSAR) can generate a 360$^\circ$ image of its surrounding environment using the collected data from a single moving track. Due to its non-linear track, the Back-Projection Algorithm (BPA) is commonly used to generate SAR images in ROSAR. Despite its superior imaging performance, BPA suffers from high computation complexity, restricting its application in real-time systems. In this paper, we propose an efficient imaging method based on robust sparse array synthesis. It first conducts range-dimension matched filtering, followed by azimuth-dimension matched filtering using a selected sparse aperture and filtering weights. The aperture and weights are computed offline in advance to ensure robustness to array manifold errors induced by the imperfect radar rotation. We introduce robust constraints on the main-lobe and sidelobe levels of filter design. The resultant robust sparse array synthesis problem is a non-convex optimization problem with quadratic constraints. An algorithm based on feasible point pursuit and successive convex approximation is devised to solve the optimization problem. Extensive simulation study and experimental evaluations using a real-world hardware platform demonstrate that the proposed algorithm can achieve image quality comparable to that of BPA, but with a substantial reduction in computational time up to 90%.

Pi-ViMo: Physiology-inspired Robust Vital Sign Monitoring using mmWave Radars

Continuous monitoring of human vital signs using non-contact mmWave radars is attractive due to their ability to penetrate garments and operate under different lighting conditions. Unfortunately, most prior research requires subjects to stay at a fixed distance from radar sensors and to remain still during monitoring. These restrictions limit the applications of radar vital sign monitoring in real life scenarios. In this paper, we address these limitations and present "Pi-ViMo", a non-contact Physiology-inspired Robust Vital Sign Monitoring system, using mmWave radars. We first derive a multi-scattering point model for the human body, and introduce a coherent combining of multiple scatterings to enhance the quality of estimated chest-wall movements. It enables vital sign estimations of subjects at any location in a radar's field of view. We then propose a template matching method to extract human vital signs by adopting physical models of respiration and cardiac activities. The proposed method is capable to separate respiration and heartbeat in the presence of micro-level random body movements (RBM) when a subject is at any location within the field of view of a radar. Experiments in a radar testbed show average respiration rate errors of 6% and heart rate errors of 11.9% for the stationary subjects and average errors of 13.5% for respiration rate and 13.6% for heart rate for subjects under different RBMs.

VALERIAN: Invariant Feature Learning for IMU Sensor-based Human Activity Recognition in the Wild

Deep neural network models for IMU sensor-based human activity recognition (HAR) that are trained from controlled, well-curated datasets suffer from poor generalizability in practical deployments. However, data collected from naturalistic settings often contains significant label noise. In this work, we examine two in-the-wild HAR datasets and DivideMix, a state-of-the-art learning with noise labels (LNL) method to understand the extent and impacts of noisy labels in training data. Our empirical analysis reveals that the substantial domain gaps among diverse subjects cause LNL methods to violate a key underlying assumption, namely, neural networks tend to fit simpler (and thus clean) data in early training epochs. Motivated by the insights, we design VALERIAN, an invariant feature learning method for in-the-wild wearable sensor-based HAR. By training a multi-task model with separate task-specific layers for each subject, VALERIAN allows noisy labels to be dealt with individually while benefiting from shared feature representation across subjects. We evaluated VALERIAN on four datasets, two collected in a controlled environment and two in the wild.

Deep Learning-Driven Edge Video Analytics: A Survey

Video, as a key driver in the global explosion of digital information, can create tremendous benefits for human society. Governments and enterprises are deploying innumerable cameras for a variety of applications, e.g., law enforcement, emergency management, traffic control, and security surveillance, all facilitated by video analytics (VA). This trend is spurred by the rapid advancement of deep learning (DL), which enables more precise models for object classification, detection, and tracking. Meanwhile, with the proliferation of Internet-connected devices, massive amounts of data are generated daily, overwhelming the cloud. Edge computing, an emerging paradigm that moves workloads and services from the network core to the network edge, has been widely recognized as a promising solution. The resulting new intersection, edge video analytics (EVA), begins to attract widespread attention. Nevertheless, only a few loosely-related surveys exist on this topic. A dedicated venue for collecting and summarizing the latest advances of EVA is highly desired by the community. Besides, the basic concepts of EVA (e.g., definition, architectures, etc.) are ambiguous and neglected by these surveys due to the rapid development of this domain. A thorough clarification is needed to facilitate a consensus on these concepts. To fill in these gaps, we conduct a comprehensive survey of the recent efforts on EVA. In this paper, we first review the fundamentals of edge computing, followed by an overview of VA. The EVA system and its enabling techniques are discussed next. In addition, we introduce prevalent frameworks and datasets to aid future researchers in the development of EVA systems. Finally, we discuss existing challenges and foresee future research directions. We believe this survey will help readers comprehend the relationship between VA and edge computing, and spark new ideas on EVA.

AttTrack: Online Deep Attention Transfer for Multi-object Tracking

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.

Individualizing Head-Related Transfer Functions for Binaural Acoustic Applications

A Head Related Transfer Function (HRTF) characterizes how a human ear receives sounds from a point in space, and depends on the shapes of one's head, pinna, and torso. Accurate estimations of HRTFs for human subjects are crucial in enabling binaural acoustic applications such as sound localization and 3D sound spatialization. Unfortunately, conventional approaches for HRTF estimation rely on specialized devices or lengthy measurement processes. This work proposes a novel lightweight method for HRTF individualization that can be implemented using commercial-off-the-shelf components and performed by average users in home settings. The proposed method has two key components: a generative neural network model that can be individualized to predict HRTFs of new subjects from sparse measurements, and a lightweight measurement procedure that collects HRTF data from spatial locations. Extensive experiments using a public dataset and in house measurement data from 10 subjects of different ages and genders, show that the individualized models significantly outperform a baseline model in the accuracy of predicted HRTFs. To further demonstrate the advantages of individualized HRTFs, we implement two prototype applications for binaural localization and acoustic spatialization. We find that the performance of a localization model is improved by 15 degree after trained with individualized HRTFs. Furthermore, in hearing tests, the success rate of correctly identifying the azimuth direction of incoming sounds increases by 183% after individualization.

CROMOSim: A Deep Learning-based Cross-modality Inertial Measurement Simulator

With the prevalence of wearable devices, inertial measurement unit (IMU) data has been utilized in monitoring and assessment of human mobility such as human activity recognition (HAR). Training deep neural network (DNN) models for these tasks require a large amount of labeled data, which are hard to acquire in uncontrolled environments. To mitigate the data scarcity problem, we design CROMOSim, a cross-modality sensor simulator that simulates high fidelity virtual IMU sensor data from motion capture systems or monocular RGB cameras. It utilizes a skinned multi-person linear model (SMPL) for 3D body pose and shape representations, to enable simulation from arbitrary on-body positions. A DNN model is trained to learn the functional mapping from imperfect trajectory estimations in a 3D SMPL body tri-mesh due to measurement noise, calibration errors, occlusion and other modeling artifacts, to IMU data. We evaluate the fidelity of CROMOSim simulated data and its utility in data augmentation on various HAR datasets. Extensive experiment results show that the proposed model achieves a 6.7% improvement over baseline methods in a HAR task.

Efficient Action Recognition Using Confidence Distillation

Modern neural networks are powerful predictive models. However, when it comes to recognizing that they may be wrong about their predictions, they perform poorly. For example, for one of the most common activation functions, the ReLU and its variants, even a well-calibrated model can produce incorrect but high confidence predictions. In the related task of action recognition, most current classification methods are based on clip-level classifiers that densely sample a given video for non-overlapping, same-sized clips and aggregate the results using an aggregation function - typically averaging - to achieve video level predictions. While this approach has shown to be effective, it is sub-optimal in recognition accuracy and has a high computational overhead. To mitigate both these issues, we propose the confidence distillation framework to teach a representation of uncertainty of the teacher to the student sampler and divide the task of full video prediction between the student and the teacher models. We conduct extensive experiments on three action recognition datasets and demonstrate that our framework achieves significant improvements in action recognition accuracy (up to 20%) and computational efficiency (more than 40%).