3D Object Detection for Self-Driving Cars Using Video and LiDAR: An Ablation Study

Methods based on 64-beam LiDAR can provide very precise 3D object detection. However, highly accurate LiDAR sensors are extremely costly: a 64-beam model can cost approximately USD 75,000. We previously proposed SLS-Fusion (sparse LiDAR and stereo fusion) to fuse low-cost four-beam LiDAR with stereo cameras that outperform most advanced stereo-LiDAR fusion methods. In this paper, and according to the number of LiDAR beams used, we analyzed how the stereo and LiDAR sensors contributed to the performance of the SLS-Fusion model for 3D object detection. Data coming from the stereo camera play a significant role in the fusion model. However, it is necessary to quantify this contribution and identify the variations in such a contribution with respect to the number of LiDAR beams used inside the model. Thus, to evaluate the roles of the parts of the SLS-Fusion network that represent LiDAR and stereo camera architectures, we propose dividing the model into two independent decoder networks. The results of this study show that-starting from four beams-increasing the number of LiDAR beams has no significant impact on the SLS-Fusion performance. The presented results can guide the design decisions by practitioners.

BERT for Activity Recognition Using Sequences of Skeleton Features and Data Augmentation with GAN

Recently, the scientific community has placed great emphasis on the recognition of human activity, especially in the area of health and care for the elderly. There are already practical applications of activity recognition and unusual conditions that use body sensors such as wrist-worn devices or neck pendants. These relatively simple devices may be prone to errors, might be uncomfortable to wear, might be forgotten or not worn, and are unable to detect more subtle conditions such as incorrect postures. Therefore, other proposed methods are based on the use of images and videos to carry out human activity recognition, even in open spaces and with multiple people. However, the resulting increase in the size and complexity involved when using image data requires the use of the most recent advanced machine learning and deep learning techniques. This paper presents an approach based on deep learning with attention to the recognition of activities from multiple frames. Feature extraction is performed by estimating the pose of the human skeleton, and classification is performed using a neural network based on Bidirectional Encoder Representation of Transformers (BERT). This algorithm was trained with the UP-Fall public dataset, generating more balanced artificial data with a Generative Adversarial Neural network (GAN), and evaluated with real data, outperforming the results of other activity recognition methods using the same dataset.

Federated learning enables big data for rare cancer boundary detection

Although machine learning (ML) has shown promise across disciplines, out-of-sample generalizability is concerning. This is currently addressed by sharing multi-site data, but such centralization is challenging/infeasible to scale due to various limitations. Federated ML (FL) provides an alternative paradigm for accurate and generalizable ML, by only sharing numerical model updates. Here we present the largest FL study to-date, involving data from 71 sites across 6 continents, to generate an automatic tumor boundary detector for the rare disease of glioblastoma, reporting the largest such dataset in the literature (n = 6, 314). We demonstrate a 33% delineation improvement for the surgically targetable tumor, and 23% for the complete tumor extent, over a publicly trained model. We anticipate our study to: 1) enable more healthcare studies informed by large diverse data, ensuring meaningful results for rare diseases and underrepresented populations, 2) facilitate further analyses for glioblastoma by releasing our consensus model, and 3) demonstrate the FL effectiveness at such scale and task-complexity as a paradigm shift for multi-site collaborations, alleviating the need for data-sharing.

Human Activity Recognition by Sequences of Skeleton Features

In recent years, much effort has been devoted to the development of applications capable of detecting different types of human activity. In this field, fall detection is particularly relevant, especially for the elderly. On the one hand, some applications use wearable sensors that are integrated into cell phones, necklaces or smart bracelets to detect sudden movements of the person wearing the device. The main drawback of these types of systems is that these devices must be placed on a person’s body. This is a major drawback because they can be uncomfortable, in addition to the fact that these systems cannot be implemented in open spaces and with unfamiliar people. In contrast, other approaches perform activity recognition from video camera images, which have many advantages over the previous ones since the user is not required to wear the sensors. As a result, these applications can be implemented in open spaces and with unknown people. This paper presents a vision-based algorithm for activity recognition. The main contribution of this work is to use human skeleton pose estimation as a feature extraction method for activity detection in video camera images. The use of this method allows the detection of multiple people’s activities in the same scene. The algorithm is also capable of classifying multi-frame activities, precisely for those that need more than one frame to be detected. The method is evaluated with the public UP-FALL dataset and compared to similar algorithms using the same dataset.

Facial Expression Recognition of Instructor Using Deep Features and Extreme Learning Machine

Classroom communication involves teacher's behavior and student's responses. Extensive research has been done on the analysis of student's facial expressions, but the impact of instructor's facial expressions is yet an unexplored area of research. Facial expression recognition has the potential to predict the impact of teacher's emotions in a classroom environment. Intelligent assessment of instructor behavior during lecture delivery not only might improve the learning environment but also could save time and resources utilized in manual assessment strategies. To address the issue of manual assessment, we propose an instructor's facial expression recognition approach within a classroom using a feedforward learning model. First, the face is detected from the acquired lecture videos and key frames are selected, discarding all the redundant frames for effective high-level feature extraction. Then, deep features are extracted using multiple convolution neural networks along with parameter tuning which are then fed to a classifier. For fast learning and good generalization of the algorithm, a regularized extreme learning machine (RELM) classifier is employed which classifies five different expressions of the instructor within the classroom. Experiments are conducted on a newly created instructor's facial expression dataset in classroom environments plus three benchmark facial datasets, i.e., Cohn-Kanade, the Japanese Female Facial Expression (JAFFE) dataset, and the Facial Expression Recognition 2013 (FER2013) dataset. Furthermore, the proposed method is compared with state-of-the-art techniques, traditional classifiers, and convolutional neural models. Experimentation results indicate significant performance gain on parameters such as accuracy, F1-score, and recall.

A Unified Deep Framework for Joint 3D Pose Estimation and Action Recognition from a Single RGB Camera

We present a deep learning-based multitask framework for joint 3D human pose estimation and action recognition from RGB sensors using simple cameras. The approach proceeds along two stages. In the first, a real-time 2D pose detector is run to determine the precise pixel location of important keypoints of the human body. A two-stream deep neural network is then designed and trained to map detected 2D keypoints into 3D poses. In the second stage, the Efficient Neural Architecture Search (ENAS) algorithm is deployed to find an optimal network architecture that is used for modeling the spatio-temporal evolution of the estimated 3D poses via an image-based intermediate representation and performing action recognition. Experiments on Human3.6M, MSR Action3D and SBU Kinect Interaction datasets verify the effectiveness of the proposed method on the targeted tasks. Moreover, we show that the method requires a low computational budget for training and inference. In particular, the experimental results show that by using a monocular RGB sensor, we can develop a 3D pose estimation and human action recognition approach that reaches the performance of RGB-depth sensors. This opens up many opportunities for leveraging RGB cameras (which are much cheaper than depth cameras and extensively deployed in private and public places) to build intelligent recognition systems.

Spatio⁻Temporal Image Representation of 3D Skeletal Movements for View-Invariant Action Recognition with Deep Convolutional Neural Networks

Designing motion representations for 3D human action recognition from skeleton sequences is an important yet challenging task. An effective representation should be robust to noise, invariant to viewpoint changes and result in a good performance with low-computational demand. Two main challenges in this task include how to efficiently represent spatio–temporal patterns of skeletal movements and how to learn their discriminative features for classification tasks. This paper presents a novel skeleton-based representation and a deep learning framework for 3D action recognition using RGB-D sensors. We propose to build an action map called SPMF (Skeleton Posture-Motion Feature), which is a compact image representation built from skeleton poses and their motions. An Adaptive Histogram Equalization (AHE) algorithm is then applied on the SPMF to enhance their local patterns and form an enhanced action map, namely Enhanced-SPMF. For learning and classification tasks, we exploit Deep Convolutional Neural Networks based on the DenseNet architecture to learn directly an end-to-end mapping between input skeleton sequences and their action labels via the Enhanced-SPMFs. The proposed method is evaluated on four challenging benchmark datasets, including both individual actions, interactions, multiview and large-scale datasets. The experimental results demonstrate that the proposed method outperforms previous state-of-the-art approaches on all benchmark tasks, whilst requiring low computational time for training and inference.

Structural Laplacian Eigenmaps for modeling sets of multivariate sequences

A novel embedding-based dimensionality reduction approach, called structural Laplacian Eigenmaps, is proposed to learn models representing any concept that can be defined by a set of multivariate sequences. This approach relies on the expression of the intrinsic structure of the multivariate sequences in the form of structural constraints, which are imposed on dimensionality reduction process to generate a compact and data-driven manifold in a low dimensional space. This manifold is a mathematical representation of the intrinsic nature of the concept of interest regardless of the stylistic variability found in its instances. In addition, this approach is extended to model jointly several related concepts within a unified representation creating a continuous space between concept manifolds. Since a generated manifold encodes the unique characteristic of the concept of interest, it can be employed for classification of unknown instances of concepts. Exhaustive experimental evaluation on different datasets confirms the superiority of the proposed methodology to other state-of-the-art dimensionality reduction methods. Finally, the practical value of this novel dimensionality reduction method is demonstrated in three challenging computer vision applications, i.e., view-dependent and view-independent action recognition as well as human-human interaction classification.