Cognitive Robotics最新文献

Some problems still exist in health management and application such as insufficient data, limited technology, and lack of professional evaluation methods by physicians with medical theory. In this study, an intelligent method is based on an analysis of physiological big data collected by wearable smartwatches. Firstly, physiological data such as pulse, heart rate, and blood oxygen were collected continuously from individuals by wearing smartwatches, and the data was digitally transmitted. Secondly, the transmitted data was sent to a health management platform by Narrow Band Internet of Things. Analyzing the data, physicians evaluated individual situations via an intelligent math model. Finally, the results were fed back to individuals through a smartphone APP to finish a medical diagnosis, disease prediction, or warning. The intelligent health management method and technology created via years of studies have been verified and will provide a new and effective strategy for health management.

The conventional application of deep reinforcement learning (DRL) to autonomous racing requires the agent to crash during training, thus limiting training to simulation environments. Further, many DRL approaches still exhibit high crash rates after training, making them infeasible for real-world use. This paper addresses the problem of safely training DRL agents for autonomous racing. Firstly, we present a Viability Theory-based supervisor that ensures the vehicle does not crash and remains within the friction limit while maintaining recursive feasibility. Secondly, we use the supervisor to ensure the vehicle does not crash during the training of DRL agents for high-speed racing. The evaluation in the open-source F1Tenth simulator demonstrates that our safety system can ensure the safety of a worst-case scenario planner on four test maps up to speeds of 6 m/s. Training agents to race with the supervisor significantly improves sample efficiency, requiring only 10,000 steps. Our learning formulation leads to learning more conservative, safer policies with slower lap times and a higher success rate, resulting in our method being feasible for physical vehicle racing. Enabling DRL agents to learn to race without ever crashing is a step towards using DRL on physical vehicles.

Road detection remains a captivating and crucial aspect of any form of autonomous driving. In this manuscript, we furnish a comprehensive appraisal of recent advancements in road lane detection, a fundamental component integral to autonomous driving. Despite numerous methodologies being proposed to augment accuracy while expediting speed, various hindrances, including lane marking variations, lighting fluctuations, and shadowy conditions, necessitate the establishment of dependable detection systems. Model-based and learning-based methods represent the two predominant techniques for lane detection. Model-based methods afford rapid computation speeds, while learning-based methods extend robustness amidst complexity. This paper delves into the techniques of lane detection and forecasts upcoming trends in the field. Collectively, this review offers a sturdy foundation for prospective research in the realm of road lane detection.

Robots such as drones, ground rovers, underwater vehicles and industrial robots have increased in popularity in recent years. Many sectors have benefited from this by increasing productivity while also decreasing costs and certain risks to humans. These robots can be controlled individually but are more efficient in a large group, also known as a swarm. However, an increase in the quantity and complexity of robots creates the need for an adequate control system. Reinforcement learning, an artificial intelligence paradigm, is an increasingly popular approach to control a swarm of unmanned vehicles. The quantity of reviews in the field of reinforcement learning-based swarm robotics is limited. We propose reviewing the various applications, algorithms and simulators on the subject to fill this gap. First, we present the current applications on swarm robotics with a focus on reinforcement learning control systems. Subsequently, we define important reinforcement learning terminologies, followed by a review of the current state-of-the-art in the field of swarm robotics utilizing reinforcement learning. Additionally, we review the various simulators used to train, validate and simulate swarms of unmanned vehicles. We finalize our review by discussing our findings and the possible directions for future research. Overall, our review demonstrates the potential and state-of-the-art reinforcement learning-based control systems for swarm robotics.

Aiming at the problem that the panoptic segmentation network based on coding structure can't accurately extract the detailed information of panoptic images, considering that there are some commonalities between semantic segmentation and instance segmentation tasks, this paper proposes a panoptic segmentation model with multi-feature fusion structure, which generates multi-scale fused feature maps for the panoptic segmentation network, uses the Atrous Spatial Pyramid Pooling to preferentially process the high-level features with rich context information, and then uses the cascade method to splice the low-level features to improve the panoptic segmentation performance of the model. By adding coordinate attention to the ASPP module of the corresponding branch, the perception ability of the model to the contour and instance center is enhanced.

To prevent the traditional particle swarm optimizer (PSO) from inefficient search in complex problem spaces, this paper presents a novel spread-based elite opposite swarm optimizer (SEOSO) for large scale optimization. Inspired by the dandelion seeds in nature, the seeds can randomly spread by wind and grow better for the next generation. To achieve this, the spread learning and elite opposite learning are introduced in SEOSO. In spread learning, the particles are divided into some subswarms and these subswarms can exchange the particles to get more useful information that improves the diversity of the swarm. In elite opposite learning, the opposite position of the particle is used to exclude the worse direction. The experiments are conducted on 35 benchmark functions to evaluate the performance of SEOSO in comparison with several state-of-the-art algorithms. The comparative results show the effectiveness of SEOSO in solving large scale optimization problems.

In the use of fire-fighting physics platform for fire alarm data correlation analysis, there are often problems such as too much data volume and insufficient accuracy of the analysis results. For such questions, this paper establishes a full-factor secondary mining mechanism for fire accidents based on the fire big data based on the correlation analysis algorithm and the clustering algorithm. The association algorithm is used to conduct full-factor primary mining on the fire-related factors in the data warehouse, and the common-sense accident attributes in the association rules are extracted. Then use the K-means clustering algorithm, where the cluster center is the relevant attribute in the fire accident record, and perform the second combined clustering of the accident elements to achieve in-depth information mining of all factors of the fire accident. Experimental results show that the improved full-factor deep information mining algorithm proposed in this paper can effectively filter 31.6% of meaningless mining results compared to the traditional single mining algorithm. It shows that the algorithm in this paper can more accurately dig out the relationship between data, and can provide more effective decision support for fire management and other work.

There are many tasks on image aesthetic assessment, such as aesthetic classification, scoring, score distribution prediction, and captions. Due to the distribution of the aesthetic score is unbalanced, the assessment models always output scores near the mean score. In this paper, we propose a fine-grained regression method for aesthetics score regression and combine position and channel attention mechanisms to enhance the aesthetic feature fusion. And by training the regression network separately from the classification network, we make the classification task a complement to the regression task. Besides, the researchers are used to using Mean Square Error (MSE) as the main evaluation metric which is inadequate in measuring the error of each interval. In order to fully consider the images of the various aesthetic score segments, instead of focusing on the intermediate aesthetic score segments because of the imbalance of the aesthetic datasets, we propose a new evaluation metric called Segmented Mean Square Errors (SMSE) to prove the advantages of the model. We divide the entire AADB dataset into 10 equal parts based on the aesthetic scores and the experiments were carried out on each of the segmented AADB datasets. In this way, images for each aesthetic score segment are fairly considered. The experimental results reveal that our method outperforms all the state-of-the-art methods on both MSE and SMSE. The dual attention modules of position and channel also make the activation maps more reasonable. Our methods make the aesthetic scoring go beyond laboratories to real life applications. Because computational visual aesthetics is a very interesting and challenging task in the field of computer vision, and computer vision is also one of the key areas of focus of this journal, the method proposed in this paper is closely related to the field covered by the journal.

Due to the numerous real-world applications of the crowd counting job, it has become a popular research topic. Modern crowd counting systems have a sophisticated structure and employ a filter on a big image size, making them difficult to use. Because these technologies are computationally intensive and difficult to implement in small surveillance systems, they are not appropriate for use in small surveillance systems. They also function poorly in a variety of sizes and densities, as well. Transfer learning and deep convolutional neural network architecture are used to create a modest but efficient network, which we describe herein. We named the proposed crowd counting architecture deep crowd transfer network (DCTNets) since it incorporates both deep learning and transfer learning principles into a single system. Among DCTNets’ key components are a detection module that is based on mask R-CNNs and an estimate module that is based on deep convolutional neural networks. In the first step, we apply transfer learning to the Mask R-CNN model using the datasets ShanghaiTech, JHU-CROWD++, and UCF-QNRF. After that, we train and evaluate the complete architecture on these datasets using the transfer learning results. Input images are sent through a Mask R-CNN, which counts individuals and segments the counted region, then through an estimation network, which estimates the population size, and finally through a merge of the outputs from the two models. According to the findings of comparative tests, the proposed model outperforms existing state-of-the-art approaches on the ShanghaiTech, JHU-CROWD++, and UCF-QNRF datasets.