Deep neural networks (DNNs) have been widely applied in mobile intelligent applications. However, their high computational complexity poses significant challenges for resource-constrained mobile devices. To address this issue, this paper proposes a multi-uncrewed aerial vehicle (UAV)-assisted mobile edge computing architecture tailored for DNN inference tasks. By hierarchically partitioning the DNN model and distributing different sub-tasks between local devices and aerial servers for collaborative processing, the system effectively reduces the computational burden on user terminals. Taking into account the factors such as unbalanced network load and limited UAV energy, a task migration mechanism is introduced to support resource coordination and load balancing among multiple UAVs. The aim is to minimize total weighted energy consumption through joint optimization of user-UAV association, DNN partitioning, UAV trajectory, task migration, and computing resource allocation. Due to the dynamic and complex nature of the resulting optimization problem, we model it as a Markov decision process, and a soft actor-critic with prioritized experience replay (SAC-PER) is proposed to solve it. Furthermore, we integrate convex optimization techniques into SAC-PER as a subroutine to allocate computing resources to enhance the learning efficiency. Simulation results show that the proposed method achieves faster convergence and reduces the total weighted energy consumption by up to 14.6% compared with baseline methods.

Research on brain-computer interfaces (BCI) has shown a great deal of interest in the use of electroencephalography (EEG) in emotion recognition. However, characteristics must be retrieved from EEG data in terms of multi-view in order to create an efficient model for emotion recognition. We propose a novel network, MDA-ML, which incorporates a multi-source domain adaptation network and uses multi-modal data in multi-view representation to address the issues of multi-feature interaction and domain adaptation. Our method improves model generalization by constructing a domain adaptation network with multi-view learning.To determine the links between features from different viewpoints and modalities, the network uses multi-feature fusion and various architectures for multi-view feature extraction.Four public emotion datasets with a variety of emotion classification tasks are used to validate our model. MDA-ML outperforms the earlier approaches in model generalization with minimal data in leave-one-subject-out tests.

Research on brain-computer interfaces (BCI) has shown a great deal of interest in the use of electroencephalography (EEG) in emotion recognition. However, characteristics must be retrieved from EEG data in terms of multi-view in order to create an efficient model for emotion recognition. We propose a novel network, MDA-ML, which incorporates a multisource domain adaptation network and uses multi-modal data in multi-view representation to address the issues of multi-feature interaction and domain adaptation. Our method improves model generalization by constructing a domain adaptation network with multi-view learning. To determine the links between features from different viewpoints and modalities, the network uses multi-feature fusion and various architectures for multi-view feature extraction. Four public emotion datasets with a variety of emotion classification tasks are used to validate our model. MDA-ML outperforms the earlier approaches in model generalization with minimal data in leave-one-subject-out tests.

Millimeter wave (mmWave) has been claimed as the viable solution for high-bandwidth vehicular communications in 5G and beyond. To realize applications in future vehicular communications, it is important to take a robust mmWave vehicular network into consideration. However, one challenge in such a network is that mmWave should provide an ultra-fast and high-rate data exchange among vehicles or vehicle-to-infrastructure (V2I). Moreover, traditional real-time channel estimation strategies are unavailable because vehicle mobility leads to a fast variation mmWave channel. To overcome these issues, a channel estimation approach for mmWave V2I communications is proposed in this paper. Specifically, by considering a fast-moving vehicle secnario, a corresponding mathematical model for a fast time-varying channel is first established. Then, the temporal variation rule between the base station and each mobile user and the determined direction-of-arrival are used to predict the time-varying channel prior information (PI). Finally, by exploiting the PI and the characteristics of the channel, the time-varying channel is estimated. The simulation results show that the scheme in this paper outperforms traditional ones in both normalized mean square error and sum-rate performance in the mmWave time-varying vehicular system.

Modern automated warehouses are equipped with one or many expensive and sophisticated equipment, such as palletizing robots, automated guided vehicles as well as an automated storage and retrieval system (AS/RS). These equipment are operated manually at many levels. These manual interruptions are accompanied by disadvantages of slow storage and retrieval speed, high operating costs and high frequency of errors in the operations. This paper presents an approach for efficient robotic stowing of items for inventory replenishment in a storage system. The objective is to enable a robotic arm system to stow items into a storage bin system and automatically generate a file to indicate which bin each object is stowed to. This would require a robust object recognition imbued with recognition history such that a previously recognized object is remembered as being stowed, even if it has been obscured by other objects subsequently during the task. A feature confidence aggregation strategy has been implemented to analyze a sequence of images containing a number of objects that are added to the storage system sequentially. The strategy is based on a weighted aggregation of ranked machine-learned classification scores and feature-matching recognition scores. This method is able to produce a high recognition rate and has been applied in the Amazon Robotic Challenge 2017 by Team Nanyang.

In order to build a stable and reliable system for the Amazon Robotics Challenge we went through a detailed study of the performance and system requirements based on the rules and our past experience of the challenge. The challenge was to build a robot that integrates grasping, vision, motion planning, among others, to be able to pick items from a shelf to specific order boxes. This paper presents the development process including component selection, module designs, and deployment. The resulting robot system has dual 6 degrees of freedom industrial arms mounted on fixed bases, which in turn are mounted on a calibrated table. The robot works with a custom-designed top-open extendable shelf. The vision system uses multiple stereo cameras mounted on a fixed calibrated frame. Feature-based comparison and machine-learning based matching are used to identify and determine item pose. The gripper system uses suction cup and the grasping strategy is pick from the top. Error recovery strategies were also implemented to ensure robust performance. During the competition, the robot was able to pick all target items with the shortest amount of time.

In this paper we exploit normalized mutual information for the nonrigid registration of multimodal images. Rather than assuming that image statistics are spatially stationary, as often done in traditional information-theoretic methods, we take into account the spatial variability through a weighted combination of global normalized mutual information and local matching statistics. Spatial relationships are incorporated into the registration criterion by adoptively adjusting the weight according to the strength of local cues. With a continuous representation of images and Parzen window estimators, we have developed closed-form expressions of the first-order variation with respect to any general, nonparametric, infinite-dimensional deformation of the image domain. To characterize the performance of the proposed approach, synthetic phantoms, simulated MRIs, and clinical data are used in a validation study. The results suggest that the augmented normalized mutual information provides substantial improvements in terms of registration accuracy and robustness.

Gene expression data provide information on the location where certain genes are active; in order for this to be useful, such a location must be registered to an anatomical atlas. Because gene expression maps are considerably different from each other - they display the expression of different genes - and from the anatomical atlas, this problem is currently addressed either manually by trained experts, or by neglecting all image information and only using the pre-segmented boundaries. In this manuscript we concentrate on data discrepancy measures that take into account image information when this is present in both the target and template images. We exploit such "bi-lateral" structures to drive the correspondence process in regions where the intensity information is inconsistent, analogously to a "motion in- painting" task. Although no ground truth can be established, and prior information clearly plays a key role, we show that our model achieves desirable results on subjective tests validated by expert subjects.