Search

In the intelligent unmanned systems, unmanned aerial vehicle (UAV) obstacle avoidance technology is the core and primary condition. Traditional algorithms are not suitable for obstacle avoidance in complex and changeable environments based on the limited sensors on UAVs. In this article, we use an end-to-end deep reinforcement learning (DRL) algorithm to achieve the UAV autonomously avoid obstacles. For the problem of slow convergence in DRL, a Multi-Branch (MB) network structure is proposed to ensure that the algorithm can get good performance in the early stage; for non-optimal decision-making problems caused by overestimation, the Revise Q-value (RQ) algorithm is proposed to ensure that the agent can choose the optimal strategy for obstacle avoidance. According to the flying characteristics of the rotor UAV, we build a V-Rep 3D physical simulation environment to test the obstacle avoidance performance. And experiments show that the improved algorithm can accelerate the convergence speed of agent and the average return of the round is increased by 25%.

Object detection plays an important role in underground intelligent vehicles and intelligent transportation systems. Due to the uneven light in underground mining scenarios, infrared cameras are one of the typical onboard sensors for environmental perception. Although object detection has been studied for decades, it still confronts the challenge of detecting infrared objects in underground mines. The contributing factors include weak and small objects in infrared images and similar environments in mining scenarios. In this paper, a Feature Enhancement and Guided Network (FEGNet) is proposed to address these problems. Based on the characteristics of infrared images, the feature enhancement module (FEM) preserves the image details from global and local perspectives to improve the discrimination of weak and small objects. To tackle the problem of overfitting caused by similar environments, a receptive-field-guided (RFG) backbone is proposed to learn multi-scale context and spatial information. The experimental results on the underground mining (UM) dataset demonstrate that the mAP of the proposed FEGNet achieves 86.1%, which is 4.6% higher than the state-of-the-art CNN-based network YOLOv7.

With the extending use of LiDAR SLAM in various areas, the interference of external disturbances on SLAM is becoming more and more obvious. Huge efforts have been made to reduce the drift error of LiDAR SLAM using graph-based methods. However, the mapping results can be severely affected by external disturbances under extreme conditions, which will limit the performance of graph-based methods. This study proposes a new strategy to reduce the static drift on a local scale by identifying and compensating the influence of external disturbances based on the localization results of LiDAR SLAM. Contrast experiments were first designed and performed to analyze the potential inducing factors of static drift, such as environment and vibration. The Kalman filter was adopted to estimate the speed and acceleration parameters based on the localization results of LiDAR SLAM. Then, an estimation criterion of static drift was established according to the interference of external disturbances on speed and acceleration. Finally, a static drift compensation method for LiDAR SLAM was proposed to compensate the drift of the pose. In the verification experiment, for 1866 data points, the identification accuracy of static drift was 97.32%, and the final positioning error of LiDAR SLAM was reduced from 4.9464 m to 0.1741 m after the compensation of static drift.

With widespread utilization of graphene oxides (GO) in industry, it is ineluctably discharged into environment and has been found to be toxic to cells and animals. Effective techniques and functional materials are necessary for the elimination of GO from wastewater. In this work, the GO was removed by the magnetic Fe3O4/MgAl-layered double hydroxide (MLDH) and its oxide (MLDO) from aqueous solution. The effects of dosage, time and solution pH on the removal capacity of the MLDH and MLDO was carried out in detail through batch experiments. The results showed that the MLDH and MLDO can remove GO from aqueous solution quickly and efficiently. The maximum adsorption capacities were 82.4 mg/g of MLDH and 86.7 mg/g of MLDO. The pseudo-second-order equation was well accorded with the adsorption kinetic data. The isothermal data followed the Langmuir and Freundlich models. The interaction mechanisms of GO were ascribed to the surface complexation and electrostatic attraction by using zeta potential determination and XRD spectra. Moreover, the MLDH and MLDO after adsorbing GO can be separated in 10 s by a magnet. The short adsorption time, excellent removal capacity and extremely fast solid-liquid separation indicated that the MLDH and MLDO were potential magnetic adsorbents to remove the GO nanoparticles in aqueous system.

To improve the safety and efficiency of train operation, autonomous driving train have developed rapidly in recent years. Among them, the signal detection is one of the most basic functions. However, due to the small size of signal light and the complicated of the railway environment, the signal detection is still a huge problem. The existing methods, such as the approach based on Hough circle transformation, are hard to meet the practical application requirements. In this paper, a real time railway signal lights detection based on Yolov5 is introduced. And a lot of experiments were conducted to prove the effectiveness of the proposed method. The experimental results show that the proposed method achieved 0.972 for both average recall rate and average accuracy rate. Besides, the detection speed of the proposed method reached astonishing 100FPS. Overall, the detection speed and accuracy both meet the practical application requirements.

At present, drones are rapidly developing in the aviation industry and are applied to all aspects of life. However, letting drones autonomously avoid obstacles is still the focus of research by aviation scholars at this stage. However, the current automation is mostly based on human experience to determine the obstacle avoidance strategy of UAV. And the method only rely on the machine to avoid obstacle is very few. In this paper, the UAV collect visual and distance sensor information to make autonomous obstacle avoidance decision through the deep reinforcement learning algorithm, and the algorithm is tested in the v-rep simulation environment.