A Multi-Modal Egocentric Activity Recognition Approach towards Video Domain Generalization

Egocentric activity recognition is a prominent computer vision task that is based on the use of wearable cameras. Since egocentric videos are captured through the perspective of the person wearing the camera, her/his body motions severely complicate the video content, imposing several challenges. In this work we propose a novel approach for domain-generalized egocentric human activity recognition. Typical approaches use a large amount of training data, aiming to cover all possible variants of each action. Moreover, several recent approaches have attempted to handle discrepancies between domains with a variety of costly and mostly unsupervised domain adaptation methods. In our approach we show that through simple manipulation of available source domain data and with minor involvement from the target domain, we are able to produce robust models, able to adequately predict human activity in egocentric video sequences. To this end, we introduce a novel three-stream deep neural network architecture combining elements of vision transformers and residual neural networks which are trained using multi-modal data. We evaluate the proposed approach using a challenging, egocentric video dataset and demonstrate its superiority over recent, state-of-the-art research works.

Designing an Egocentric Video-Based Dashboard to Report Hand Performance Measures for Outpatient Rehabilitation of Cervical Spinal Cord Injury

Background

Functional use of the upper extremities (UEs) is a top recovery priority for individuals with cervical spinal cord injury (cSCI), but the inability to monitor recovery at home and limitations in hand function outcome measures impede optimal recovery.

Objectives

We developed a framework using wearable cameras to monitor hand use at home and aimed to identify the best way to report information to clinicians.

Methods

A dashboard was iteratively developed with clinician (n = 7) input through focus groups and interviews, creating low-fidelity prototypes based on recurring feedback until no new information emerged. Affinity diagramming was used to identify themes and subthemes from interview data. User stories were developed and mapped to specific features to create a high-fidelity prototype.

Results

Useful elements identified for a dashboard reporting hand performance included summaries to interpret graphs, a breakdown of hand posture and activity to provide context, video snippets to qualitatively view hand use at home, patient notes to understand patient satisfaction or struggles, and time series graphing of metrics to measure trends over time.

Conclusion

Involving end-users in the design process and breaking down user requirements into user stories helped identify necessary interface elements for reporting hand performance metrics to clinicians. Clinicians recognized the dashboard's potential to monitor rehabilitation progress, provide feedback on hand use, and track progress over time. Concerns were raised about the implementation into clinical practice, therefore further inquiry is needed to determine the tool's feasibility and usefulness in clinical practice for individuals with UE impairments.

Fusion of Appearance and Motion Features for Daily Activity Recognition from Egocentric Perspective

Vidos from a first-person or egocentric perspective offer a promising tool for recognizing various activities related to daily living. In the egocentric perspective, the video is obtained from a wearable camera, and this enables the capture of the person's activities in a consistent viewpoint. Recognition of activity using a wearable sensor is challenging due to various reasons, such as motion blur and large variations. The existing methods are based on extracting handcrafted features from video frames to represent the contents. These features are domain-dependent, where features that are suitable for a specific dataset may not be suitable for others. In this paper, we propose a novel solution to recognize daily living activities from a pre-segmented video clip. The pre-trained convolutional neural network (CNN) model VGG16 is used to extract visual features from sampled video frames and then aggregated by the proposed pooling scheme. The proposed solution combines appearance and motion features extracted from video frames and optical flow images, respectively. The methods of mean and max spatial pooling (MMSP) and max mean temporal pyramid (TPMM) pooling are proposed to compose the final video descriptor. The feature is applied to a linear support vector machine (SVM) to recognize the type of activities observed in the video clip. The evaluation of the proposed solution was performed on three public benchmark datasets. We performed studies to show the advantage of aggregating appearance and motion features for daily activity recognition. The results show that the proposed solution is promising for recognizing activities of daily living. Compared to several methods on three public datasets, the proposed MMSP-TPMM method produces higher classification performance in terms of accuracy (90.38% with LENA dataset, 75.37% with ADL dataset, 96.08% with FPPA dataset) and average per-class precision (AP) (58.42% with ADL dataset and 96.11% with FPPA dataset).

YOLO Series for Human Hand Action Detection and Classification from Egocentric Videos

Hand detection and classification is a very important pre-processing step in building applications based on three-dimensional (3D) hand pose estimation and hand activity recognition. To automatically limit the hand data area on egocentric vision (EV) datasets, especially to see the development and performance of the "You Only Live Once" (YOLO) network over the past seven years, we propose a study comparing the efficiency of hand detection and classification based on the YOLO-family networks. This study is based on the following problems: (1) systematizing all architectures, advantages, and disadvantages of YOLO-family networks from version (v)1 to v7; (2) preparing ground-truth data for pre-trained models and evaluation models of hand detection and classification on EV datasets (FPHAB, HOI4D, RehabHand); (3) fine-tuning the hand detection and classification model based on the YOLO-family networks, hand detection, and classification evaluation on the EV datasets. Hand detection and classification results on the YOLOv7 network and its variations were the best across all three datasets. The results of the YOLOv7-w6 network are as follows: FPHAB is P = 97% with Thesh_IOU = 0.5; HOI4D is P = 95% with Thesh_IOU = 0.5; RehabHand is larger than 95% with Thesh_IOU = 0.5; the processing speed of YOLOv7-w6 is 60 fps with a resolution of 1280 × 1280 pixels and that of YOLOv7 is 133 fps with a resolution of 640 × 640 pixels.

Measuring Hand Use in the Home after Cervical Spinal Cord Injury Using Egocentric Video

Egocentric video has recently emerged as a potential solution for monitoring hand function in individuals living with tetraplegia in the community, especially for its ability to detect functional use in the home environment. The aim of this study was to develop and validate a wearable vision-based system for measuring hand use in the home among individuals living with tetraplegia. Several deep learning algorithms for detecting functional hand-object interactions were developed and compared. The most accurate algorithm was used to extract measures of hand function from 65 h of unscripted video recorded at home by 20 participants with tetraplegia. These measures were: the percentage of interaction time over total recording time (Perc); the average duration of individual interactions (Dur); and the number of interactions per hour (Num). To demonstrate the clinical validity of the technology, egocentric measures were correlated with validated clinical assessments of hand function and independence (Graded Redefined Assessment of Strength, Sensibility and Prehension [GRASSP], Upper Extremity Motor Score [UEMS], and Spinal Cord Independent Measure [SCIM]). Hand-object interactions were automatically detected with a median F1-score of 0.80 (0.67-0.87). Our results demonstrated that higher UEMS and better prehension were related to greater time spent interacting, whereas higher SCIM and better hand sensation resulted in a higher number of interactions performed during the egocentric video recordings. For the first time, measures of hand function automatically estimated in an unconstrained environment in individuals with tetraplegia have been validated against internationally accepted measures of hand function. Future work will necessitate a formal evaluation of the reliability and responsiveness of the egocentric-based performance measures for hand use.

Egocentric Gesture Recognition Using 3D Convolutional Neural Networks for the Spatiotemporal Adaptation of Collaborative Robots

Collaborative robots are currently deployed in professional environments, in collaboration with professional human operators, helping to strike the right balance between mechanization and manual intervention in manufacturing processes required by Industry 4.0. In this paper, the contribution of gesture recognition and pose estimation to the smooth introduction of cobots into an industrial assembly line is described, with a view to performing actions in parallel with the human operators and enabling interaction between them. The proposed active vision system uses two RGB-D cameras that record different points of view of gestures and poses of the operator, to build an external perception layer for the robot that facilitates spatiotemporal adaptation, in accordance with the human's behavior. The use-case of this work is concerned with LCD TV assembly of an appliance manufacturer, comprising of two parts. The first part of the above-mentioned operation is assigned to a robot, strengthening the assembly line. The second part is assigned to a human operator. Gesture recognition, pose estimation, physical interaction, and sonic notification, create a multimodal human-robot interaction system. Five experiments are performed, to test if gesture recognition and pose estimation can reduce the cycle time and range of motion of the operator, respectively. Physical interaction is achieved using the force sensor of the cobot. Pose estimation through a skeleton-tracking algorithm provides the cobot with human pose information and makes it spatially adjustable. Sonic notification is added for the case of unexpected incidents. A real-time gesture recognition module is implemented through a Deep Learning architecture consisting of Convolutional layers, trained in an egocentric view and reducing the cycle time of the routine by almost 20%. This constitutes an added value in this work, as it affords the potential of recognizing gestures independently of the anthropometric characteristics and the background. Common metrics derived from the literature are used for the evaluation of the proposed system. The percentage of spatial adaptation of the cobot is proposed as a new KPI for a collaborative system and the opinion of the human operator is measured through a questionnaire that concerns the various affective states of the operator during the collaboration.

Braille Block Detection via Multi-Objective Optimization from an Egocentric Viewpoint

In this paper, we propose a method to detect Braille blocks from an egocentric viewpoint, which is a key part of many walking support devices for visually impaired people. Our main contribution is to cast this task as a multi-objective optimization problem and exploits both the geometric and the appearance features for detection. Specifically, two objective functions were designed under an evolutionary optimization framework with a line pair modeled as an individual (i.e., solution). Both of the objectives follow the basic characteristics of the Braille blocks, which aim to clarify the boundaries and estimate the likelihood of the Braille block surface. Our proposed method was assessed by an originally collected and annotated dataset under real scenarios. Both quantitative and qualitative experimental results show that the proposed method can detect Braille blocks under various environments. We also provide a comprehensive comparison of the detection performance with respect to different multi-objective optimization algorithms.

Toward Joint Acquisition-Annotation of Images with Egocentric Devices for a Lower-Cost Machine Learning Application to Apple Detection

Since most computer vision approaches are now driven by machine learning, the current bottleneck is the annotation of images. This time-consuming task is usually performed manually after the acquisition of images. In this article, we assess the value of various egocentric vision approaches in regard to performing joint acquisition and automatic image annotation rather than the conventional two-step process of acquisition followed by manual annotation. This approach is illustrated with apple detection in challenging field conditions. We demonstrate the possibility of high performance in automatic apple segmentation (Dice 0.85), apple counting (88 percent of probability of good detection, and 0.09 true-negative rate), and apple localization (a shift error of fewer than 3 pixels) with eye-tracking systems. This is obtained by simply applying the areas of interest captured by the egocentric devices to standard, non-supervised image segmentation. We especially stress the importance in terms of time of using such eye-tracking devices on head-mounted systems to jointly perform image acquisition and automatic annotation. A gain of time of over 10-fold by comparison with classical image acquisition followed by manual image annotation is demonstrated.

The everyday statistics of objects and their names: How word learning gets its start

A key question in early word learning is how infants learn their first object names despite a natural environment thought to provide messy data for linking object names to their referents. Using head cameras worn by 7 to 11-month-old infants in the home, we document the statistics of visual objects, spoken object names, and their co-occurrence in everyday meal time events. We show that the extremely right skewed frequency distribution of visual objects underlies word-referent co-occurrence statistics that set up a clear signal in the noise upon which infants could capitalize to learn their first object names.

Student perspectives on using egocentric video recorded by smart glasses to assess communicative and clinical skills with standardised patients

INTRODUCTION

This exploratory study evaluated student perceptions of their ability to self- and peer assess (i) interpersonal communication skills and (ii) clinical procedures (a head and neck examination) during standardised patient (SP) interactions recorded by Google Glass compared to a static camera.

METHODS

Students compared the Google Glass and static camera recordings using an instrument consisting of 20 Likert-type items and four open- and closed-text items. The Likert-type items asked students to rate how effectively they could assess specific aspects of interpersonal communication and a head and neck examination in these two different types of recordings. The interpersonal communication items included verbal, paraverbal and non-verbal subscales. The open- and closed-text items asked students to report on more globally the differences between the two types of recordings. Descriptive and inferential statistical analyses were conducted for all survey items. An inductive thematic analysis was conducted to determine qualitative emergent themes from the open-text questions.

RESULTS

Students found the Glass videos more effective for assessing verbal (t₂₂ = 2.091, P = 0.048) and paraverbal communication skills (t₂₂ = 3.304, P = 0.003), whilst they reported that the static camera video was more effective for assessing non-verbal communication skills (t₂₂ = -2.132, P = 0.044). Four principle themes emerged from the students' open-text responses comparing Glass to static camera recordings for self- and peer assessment: (1) first-person perspective, (2) assessment of non-verbal communication, (3) audiovisual experience and (4) student operation of Glass.

DISCUSSION AND CONCLUSION

Our findings suggest that students perceive that Google Glass is a valuable tool for facilitating self- and peer assessment of SP examinations because of students' perceived ability to emphasise and illustrate communicative and clinical activities from a first-person perspective.

The Developing Infant Creates a Curriculum for Statistical Learning

New efforts are using head cameras and eye-trackers worn by infants to capture everyday visual environments from the point of view of the infant learner. From this vantage point, the training sets for statistical learning develop as the sensorimotor abilities of the infant develop, yielding a series of ordered datasets for visual learning that differ in content and structure between timepoints but are highly selective at each timepoint. These changing environments may constitute a developmentally ordered curriculum that optimizes learning across many domains. Future advances in computational models will be necessary to connect the developmentally changing content and statistics of infant experience to the internal machinery that does the learning.

A Developmental Approach to Machine Learning?

Visual learning depends on both the algorithms and the training material. This essay considers the natural statistics of infant- and toddler-egocentric vision. These natural training sets for human visual object recognition are very different from the training data fed into machine vision systems. Rather than equal experiences with all kinds of things, toddlers experience extremely skewed distributions with many repeated occurrences of a very few things. And though highly variable when considered as a whole, individual views of things are experienced in a specific order - with slow, smooth visual changes moment-to-moment, and developmentally ordered transitions in scene content. We propose that the skewed, ordered, biased visual experiences of infants and toddlers are the training data that allow human learners to develop a way to recognize everything, both the pervasively present entities and the rarely encountered ones. The joint consideration of real-world statistics for learning by researchers of human and machine learning seems likely to bring advances in both disciplines.

Real-world visual statistics and infants' first-learned object names

We offer a new solution to the unsolved problem of how infants break into word learning based on the visual statistics of everyday infant-perspective scenes. Images from head camera video captured by 8 1/2 to 10 1/2 month-old infants at 147 at-home mealtime events were analysed for the objects in view. The images were found to be highly cluttered with many different objects in view. However, the frequency distribution of object categories was extremely right skewed such that a very small set of objects was pervasively present-a fact that may substantially reduce the problem of referential ambiguity. The statistical structure of objects in these infant egocentric scenes differs markedly from that in the training sets used in computational models and in experiments on statistical word-referent learning. Therefore, the results also indicate a need to re-examine current explanations of how infants break into word learning.This article is part of the themed issue 'New frontiers for statistical learning in the cognitive sciences'.

Exploring Architectural Details Through a Wearable Egocentric Vision Device

Augmented user experiences in the cultural heritage domain are in increasing demand by the new digital native tourists of 21st century. In this paper, we propose a novel solution that aims at assisting the visitor during an outdoor tour of a cultural site using the unique first person perspective of wearable cameras. In particular, the approach exploits computer vision techniques to retrieve the details by proposing a robust descriptor based on the covariance of local features. Using a lightweight wearable board, the solution can localize the user with respect to the 3D point cloud of the historical landmark and provide him with information about the details at which he is currently looking. Experimental results validate the method both in terms of accuracy and computational effort. Furthermore, user evaluation based on real-world experiments shows that the proposal is deemed effective in enriching a cultural experience.

Recognition of Activities of Daily Living with Egocentric Vision: A Review

Video-based recognition of activities of daily living (ADLs) is being used in ambient assisted living systems in order to support the independent living of older people. However, current systems based on cameras located in the environment present a number of problems, such as occlusions and a limited field of view. Recently, wearable cameras have begun to be exploited. This paper presents a review of the state of the art of egocentric vision systems for the recognition of ADLs following a hierarchical structure: motion, action and activity levels, where each level provides higher semantic information and involves a longer time frame. The current egocentric vision literature suggests that ADLs recognition is mainly driven by the objects present in the scene, especially those associated with specific tasks. However, although object-based approaches have proven popular, object recognition remains a challenge due to the intra-class variations found in unconstrained scenarios. As a consequence, the performance of current systems is far from satisfactory.