FIGURE 8 - available via license: Creative Commons Attribution 4.0 International
Content may be subject to copyright.
Source publication
![FIGURE 6. Description of Openpose architecture [25].](publication/357359945/figure/fig3/AS:1105818988621827@1640659200237/Description-of-Openpose-architecture-25_Q320.jpg)
This paper proposes a posture recognition system that can be applied for medical surveillance. The proposed method estimates human posture using mobilenetV2 and long short-term memory (LSTM) to extract the important features of an image. The output of the system was a fully estimated skeleton. We used seven human indoor postures, including lying, s...
Context in source publication
Similar publications
Collecting large datasets for investigations into human locomotion is an expensive and labor-intensive process. Methods for 3D human pose estimation in the wild are becoming increasingly accurate and could soon be sufficient to assist with the collection of datasets for analysis into running kinematics from TV broadcast data. In the domain of biome...
Citations
... ResNet-50 and MobileNetV2 were used as backbones for the Faster R-CNN and SSD detectors. The selection of these backbones was justified by the good results presented in [43][44][45][46]. The training times for the detectors were 2.25 h and 0.25 h. ...
Advances in machine learning algorithms have allowed object detection and classification to become booming areas. The detection of objects, such as rocks, in mining operations is affected by fog, snow, suspended particles, and high lighting. These environmental conditions can stop the development of mining work, which entails a considerable increase in operating costs. It is vital to select a machine learning algorithm that is accurate, fast, and contributes to lower operational costs because of the aforementioned environmental situations. In this study, the Viola-Jones algorithm, Aggregate Channel Features (ACF), Faster Regions with Convolutional Neural Networks (Faster R-CNN), Single-Shot Detector (SSD), and You Only Look Once (YOLO) version 4 were analyzed, considering the precision metrics, recall, AP50, and average detection time. In our preliminary tests, we have observed that the differences between YOLO v4 and the latest versions are not substantial forthe specific problem of rock detection addressed in our article. Therefore, YOLO v4 is an appropriate and representative choice for evaluating the effectiveness of existing methods in our study. The YOLO v4 algorithm performed the best overall, whereas the SSD algorithm performed the fastest. The results indicate that the YOLO v4 algorithm is a promising candidate for detecting rocks with visual contamination in mining operations.
... Pose estimation uses key points and body joints, such as the elbows and wrists, to locate and track human posture automatically, in addition to determining the orientation of body limbs [6]. This task has a lot of potential in many applications, which include tracking human body movement or analyzing and detecting inappropriate human behavior [7]- [9]. Pose estimation is used in sports performance analysis to track and evaluate human movement accuracy, as well as in a variety of other fields like human-computer interaction (HCI) and augmented reality (AR). ...
In the computer vision field, human action recognition depending on pose estimation recently made considerable progress, especially by using deep learning, which improves recognition performance. Therefore, it has been employed in various applications, including sports and physical activity follow-up. This paper presents a technique for recognizing the human posture in different images and matching their pose similarity. This aims to evaluate the viability of employing computer vision techniques to verify a person's body pose during exercise and determine whether the pose is executed properly. Exercise is one strategy we use to maintain our health throughout life. Gymnastics and yoga are two examples of this type of exercise. The proposed algorithm identifies human action by recognizing the body's key points. The OpenPose library has been used to detect 18 key points of the human body. The action classification task is performed using the support vector machine (SVM) algorithm. Then, the algorithm computes the similarity of the human pose by comparing a model image to a test image to determine the matching score. Evaluations show that our method can perform at a competitive or state-of-the-art performance on a number of body pose datasets.
... Human pose estimation is a pre-requisite task in many application scenarios such as Human Behavior Analysis [1,2], human-computer interaction [3], and medical rehabilitation assistance [4] and so on. The main approach is to predict the location of key points of the human body by building neural networks, so proposing models with high recognition accuracy is a hot research topic. ...
... The window size needs to be designed to be divisible by the input size. So we When is 1, the window size is (4,3), which is greater than the calculation amount of HMSA. When is 2, 4, and 8, the calculation amount is smaller than the latter. ...
To tackle the issues of numerous parameters, high computational complexity, and extended detection time prevalent in current human pose estimation network models, we have incorporated an hourglass structure to create a lightweight single-path network model, which has fewer parameters and a shorter computation time. To ensure model accuracy, we have implemented a window self-attention mechanism with a reduced parameter count. Additionally, we have redesigned this self-attention module to effectively extract local and global information, thereby enriching the feature information learned by the model. This module merges with the inverted residual network architecture, creating a separate module of WGNet. Finally, WGNet can be flexibly embedded into different stages of the model. Training and validation on COCO and MPII datasets demonstrate that this model reduces the number of parameters by 25%, computational complexity by 41%, and inference time by nearly two times, compared to Hrformer, which also utilizes the windowed self-attention mechanism, at the cost of only 3.5% accuracy.
... to be able to run on edge computing devices (e.g., the NVIDIA Jetson TX2), the MobileNetV2 architecture was used to replace the VGG model [115] to produce a lightweight version of OpenPose in some works [60], [61]. In addition, the authors in [60] attempted to enhance the performance of OpenPose by training it with a new augmented dataset, utilizing data augmentation methods such as flipping, clipping, and random scaling, to address the issue of missing fall motion samples. ...
... Data cleaning involves two main steps: removing invalid or unnecessary data and handling missing data. Skeleton data is considered invalid if it belongs to a person who is not involved in the fall event [41], [62], [46], or if it lacks specific keypoints such as the neck, hip keypoints [61], [63], or if it contains wrong keypoint data such as the detection of other objects, such as a chair, as a human body [66]. Some keypoints that do not contribute to fall detection can be removed from the whole detected keypoint sets, as argued by the authors in [43], [58], [59], [67]. ...
... In [58], 2D skeleton data extracted using OpenPose [26] were fed into a stacked 2-layer LSTM [79] with 40 LSTM cells to generate deep temporal features before connecting to the Softmax classifier as shown in Fig. 8. Additionally, to reduce the internal covariate shift problem, the authors used batch normalization (BatchNorm) [141] after the input layer and the RNN layers. Similarly, the authors in [60], [61] feed the processed skeleton data into LSTM to extract DL features for later recognition step. ...
In the elderly population, falls are one of the leading causes of fatal and non-fatal injuries. Fall detection and early alarms play an important role in mitigating the negative effects of falls, especially given the growing proportion of the elderly population. Due to their non-intrusive nature, data availability, and low deployment costs, RGB videos have been used in many previous studies to detect falls. The RGB data, however, can be affected by background environment changes, resulting in non-recognition. To overcome these challenges, many researchers propose extracting skeleton data from RGB videos and using it for fall detection. Although there have been multiple surveys on fall detection, most of them focus on assessing fall detection systems using different kinds of sensors, and a comprehensive evaluation of skeleton-based fall detection in RGB videos is lacking. In this paper, we examine the most recent advances in skeleton-based fall detection in RGB videos, from handcrafted feature-based methods to advanced deep learning algorithms. Further, we present several skeleton-based fall detection techniques and their performance results on various benchmark datasets, along with challenges and future directions in this field.
... Based on the experiment result, we recommended MobileNetV2 model to identify medicinal plant leaves according to the dataset that has been collected. MobileNetV2 was chosen for further research because it got the best accuracy and shorter computation time in recognizing the image of medicinal plant leaves [46]- [48]. www.ijacsa.thesai.org ...
Plant leaves are another common source of information for determining plant species. According to the dataset that has been collected, we propose transfer learning models VGG16, VGG19, and MobileNetV2 to examine the distinguishing features to identify medicinal plant leaves. We also improved algorithm using fine-tuning strategy and analyzed a comparison with and without a fine-tuning strategy to transfer learning models performance. Several protocols or steps were used to conduct this study, including data collection, data preparation, feature extraction, classification, and evaluation. The distribution of training and validation data is 80% for training data and 20% for validation data, with 1500 images of thirty species. The testing data consisted of a total of 43 images of 30 species. Each species class consists of 1-3 images. With a validation accuracy of 96.02 percent, MobileNetV2 with fine-tuning had the best validation accuracy. MobileNetV2 with fine-tuning also had the best testing accuracy of 81.82%.
To address the problems of a large number of parameters, high computational complexity and long detection time in many current human pose estimation network models, we adopt an hourglass structure to build a single-way lightweight network model with fewer parameters and shorter computation time. In order to ensure the accuracy of the model, we use window self-attention mechanism with fewer number of parameters, and this self-attention module that can effectively extract local and global information is redesigned on its basis to make the feature information learned by the model richer, and it merges with the inverted residual network architecture to become a separate module of WGNet (Global and Local Networks, a new structure that we construct), and finally WGNet is flexibly embedded into different stages of the model. The results of training and validation on COCO (Common Objects in Context) and MPII (Max Planck Institute for Informatics) datasets show that the model reduces the number of parameters by 25%, the computational complexity by 41%, and the inference time by nearly two times compared with HRFormer, which also uses the windowed self-attention mechanism, at the expense of only 3.5% accuracy.
Machine-learning based human posture-prediction tools can potentially be robust alternatives to motion capture measurements. Existing posture-prediction approaches are confined to two-handed load-handling activities performed at heights below 120 cm from the floor and to predicting a limited number of body-joint coordinates/angles. Moreover, the extrapolating power of these tools beyond the range of the input dataset they were trained for (e.g., for underweight, overweight, or left-handed individuals) has not been investigated. In this study, we trained/validated/tested two posture-prediction (for full-body joint coordinates and angles) artificial neural networks (ANNs) using both 70%/15%/15% random-hold-out and leave-one-subject-out methods, based on a comprehensive kinematic dataset of forty-one full-body skin markers collected from twenty right-handed normal-weight (BMI = 18–26 kg/m2) subjects. Subjects performed 204 one- and two-handed unloaded activities at different vertical (0 to 180 cm from the floor) and horizontal (up to 60 cm lateral and/or anterior) destinations. Subsequently, the extrapolation capability of the trained/validated/tested ANNs was evaluated using data collected from fifteen additional subjects (unseen by the ANNs); three individuals in five groups: underweight, overweight, obese, left-handed individuals, and subjects with a hand-load. Results indicated that the ANNs predicted body joint coordinates and angles during various activities with errors of ∼ 25 mm and ∼ 10°, respectively; considerable improvements when compared to previous posture-prediction ANNs. Extrapolation errors of our ANNs generally remained within the error range of existing ANNs with obesity and being left-handed having, respectively, the most and least compromising effects on their accuracy. These easy-to-use ANNs appear, therefore, to be robust alternatives to common posture-measurement approaches.
