Search

In recent years, 3D object detection based on LiDAR point clouds is a key component of autonomous driving. In pursuit of enhancing the accuracy of 3D point cloud feature extraction and point cloud detection, this paper introduces a novel 3D object detection model, termed as Graph Self-Attention-RCNN (GA-RCNN). This model is designed to integrate voxel information and point location information, enhancing the quality of 3D object proposals while maintaining contextual accuracy. The first stage rectifies the previous approach that relied on local features for preselected boxes, overlooking crucial global contextual information. An improved method is suggested in this work, utilizing BEV to capture long-range dependencies via a cross-attention mechanism. The second stage addresses the overreliance on local neighborhood point feature extraction. The Graph Self-Attention Pooling method is proposed, characterized by its dynamic computation of contribution weights for inputs. This enhances the model's flexibility and generalization performance. Extensive evaluations on KITTI and Waymo datasets demonstrate GA-RCNN's superior accuracy compared to other methods, affirming its efficacy in 3D object detection. [ABSTRACT FROM AUTHOR]

How to learn informative representations from Electromyography (EMG) signals is of vital importance for myoelectric control systems. Traditionally, hand-crafted features are extracted from individual EMG channels and combined together for pattern recognition. The spatial topological information between different channels can also be informative, which is seldom considered. This paper presents a radically novel approach to extract spatial structural information within diverse EMG channels based on the symmetric positive definite (SPD) manifold. The object is to learn non-Euclidean representations inside EMG signals for myoelectric pattern recognition. The performance is compared with two classical feature sets using accuracy and F1-score. The algorithm is tested on eleven gestures collected from ten subjects, and the best accuracy reaches 84.85%±5.15% with an improvement of 4.04%~20.25%, which outperforms the contrast method, and reaches a significant improvement with the Wilcoxon signed-rank test. Eleven gestures from three public databases involving Ninapro DB2, DB4, and DB5 are also evaluated, and better performance is observed. Furthermore, the computational cost is less than the contrast method, making it more suitable for low-cost systems. It shows the effectiveness of the presented approach and contributes a new way for myoelectric pattern recognition. [ABSTRACT FROM AUTHOR]

Soft exoskeletons have demonstrated the potential to save energy, but their efficiency is sensitive to variations in human gait cadence. This work aims to develop adaptive gait state detection and iterative force control methods for a soft exoskeleton to reduce human walking metabolic cost consistently, while the user may change walking cadence. The proposed approach is motivated by the rhythmicity of gait and applies an iterative learning concept to enhance the exoskeleton’s adaptability to varying walking conditions. The gait state detection method proposed for the designed exoskeleton combines two feature extraction algorithms, which can learn from the present and past body kinematic data, to provide accurate user gait state detection. Based on the state, the proposed force control method iteratively adjusts the commands to keep track of the desired profile. Experiments have been conducted on healthy subjects walking with varying cadence using the soft exoskeleton. Promising results were presented in separate validation tests. Moreover, metabolic costs of subjects walking under one unpowered and two powered conditions, where the assistance profiles were produced by classical methods and the proposed methods, showed that the proposed methods can effectively improve the exoskeleton’s ability to save human energy of walking with varying cadence. Note to Practitioners—Lower limb exoskeletons have demonstrated the potential to save human energy in medical and industrial applications. The main purpose of this work is to solve the exoskeleton assistance efficiency loss problem for users walking with changing cadence. Constant cadence is unlikely maintained during natural human walking. Few existing exoskeletons could retain high efficiency under user cadence changes, limited by their control system capability. This work presents a new cable-driven cadence-insensitive soft exoskeleton, which is purposely designed with two adaptive methods to enable the device to offer consistent benefit to users walking with varying cadence. The proposed methods are inspired by the rhythmicity of human gait and can be iteratively reconfigured to perform accurate human gait state detection and assistive force tracking. The proposed methods have the potential to be integrated into other human-oriented robots to improve their adaptability. This work can greatly enhance the possibility of using the walking assist robotic devices in more practical applications. [ABSTRACT FROM AUTHOR]

In sEMG-based recognition systems, accuracy is severely worsened by disturbances, such as electrode shifts by doffing/donning. Traditional recognition models are fixed or static, with limited abilities to work in the presence of the disturbances. In this paper, a transfer learning method is proposed to reduce the impact of electrode shifts. In the proposed method, a novel activation angle is introduced to locate electrodes within a polar coordinate system. An adaptive transformation is utilized to correct electrode-shifted sEMG samples. The transformation is based on estimated shifts relative to the initial position. The experiments acquisition data from ten subjects consist of sEMG signals under eight gestures in seven or nine arbitrary positions, and recorded shifts from a 3D-printed annular ruler. In our extensive experiments, the errors between recorded shifts (as the reference) and estimated shifts is about −0.017 ± 0.13 radians. Eight gestures recognition results have shown an average accuracy around 79.32%, which represents a significant improvement over the 35.72% (p < 0.0001) average accuracy of results obtained using nonadaptive models, and 60.99% (p < 0.0001) results of the other method iGLCM (an improved gray-level co-occurrence matrix). More importantly, by only using one-label samples, the proposed method updates the pre-trained model in an initial position. As a result, the pre-trained model can be adaptively corrected to recognize eight-label gestures in arbitrarily rotary positions. It is proven a highly efficient way to relieve subjects’ re-training burden of sEMG-based rehabilitation systems. [ABSTRACT FROM AUTHOR]

Missing data are a common drawback that pattern recognition techniques need to handle when solving real-life classification tasks. This paper first discusses problems in handling high-dimensional samples with missing values by the Gaussian mixture model (GMM). Since fitting the GMM by directly using high-dimensional samples as inputs is difficult due to the convergence and stability issues, a novel method is proposed to build the high-dimensional GMM by extending a reduced-dimensional GMM to the full-dimensional space. Based on the extended full-dimensional GMM, two approaches, namely, marginalization and conditional-mean imputation, are proposed to classify samples with missing data in online phase. Then, the proposed methods were employed to recognize hand motions from surface electromyography (sEMG) signals, and more than 75% of classification accuracy of motions can be obtained even if 50% of sEMG signals were missing. Comparisons with normal mean and zero imputations also demonstrate the improvements of the proposed methods. Finally, a control scheme for a myoelectric hand was designed by involving the novel methods, and online experiments confirm the ability of the proposed methods to improve the safety and stability of practical systems. [ABSTRACT FROM PUBLISHER]