Search

Twin delayed deep deterministic (TD3) policy gradient has several limitations when applied in planning a path in environment with a number of dilemmas according to our experiment, due to the complexity of the robot path planning task, the rate of convergence of TD3 algorithm is slow and the rate of collision is high. To address this problem, deep dense dueling twin delayed deep deterministic (D3-TD3) architecture is proposed, a method that preserves important information from cross-layer inputs through dense connections and divides the network into a value function and a dominance function, thus, allowing for faster convergence when solving complex tasks. Finally, a spatial model based on three-dimension (3D) point cloud is built, and simulation experimental results show that in static environment, the algorithm proposed in the paper has 40.6% fewer collisions compared to TD3, 30% fewer collisions compared to TD3-BC, 19.2% fewer collisions compared to Dueling TD3 and 17.4% fewer collisions compared to deep dense TD3. In dynamic and static environment, the algorithm proposed in the paper has 34.4% fewer collisions compared to TD3, 24% fewer collisions compared to TD3-BC, 6% fewer collisions compared to Dueling TD3 and 25% fewer collisions compared to deep dense TD3. [ABSTRACT FROM AUTHOR]

In order to solve the problem of environment generalization in continuous state space, an obstacle avoidance method based on region location is proposed. The method is divided into three steps: (1) Using Region Proposal Network (RPN) to localize the obstacle area; (2). The environment map is established by the regional position mapping relation; and (3). The Deep Q-Learning Network (DQN) is used to realize collision detection of the robot, then pixel collision detection module is introduced and finally the pixel collision simulation distance sensor is combined to obtain the distance between the robot and obstacle and whether the collision or not. In this paper, the experiments were carried out in static obstacle environment and in dynamic and static obstacle environment for robot obstacle avoidance tasks. Experimental results show that the problem of environment generalization can be effectively solved by introducing pixel collision detection in the process of robot obstacle avoidance, and the network model trained in a dynamic environment has some generalization ability. [ABSTRACT FROM AUTHOR]

In order to solve the problem of difficult and low precision in the detection of safety helmet wearing in the complex pose of construction worker, a detection method of safety helmet wearing based on pose estimation is proposed. In the pose estimation model of OpenPose, the residual network optimized feature extraction is introduced to obtain the skeletal point information of the construction worker, and then the pose of the construction worker is estimated based on the skeletal point information, three‐point localization method is proposed for the front and back pose, and skin colour detection method is proposed for the side pose, and then to determine the head region. The YOLO v4 is used to detect the safety helmet region, and then the construction worker's safety helmet wearing is judged according to whether the head region intersects the safety helmet region or not. Experimental results show that the detection accuracy of the method is higher than other methods, and the adaptability to the environment is stronger. [ABSTRACT FROM AUTHOR]

The one-dimensional Otsu method is an adaptive threshold method. It obtains the optimal threshold for image segmentation by the maximum between-class variance, without considering the minimum within-class variance. As the background of water surface image is mostly uniform, using this feature, the threshold selection tactics adopt the combination of the one-dimensional Otsu method and the uniformity measurement, proposes the threshold segmentation method based on uniformity measurement, and adopts the performance evaluation method based on GT image to compare the segmentation result. Experimental results demonstrate that effectiveness of the improved Otsu method is generally better than the traditional Otsu method, and the other four commonly used threshold segmentation methods for the water surface image, which improves the segmentation accuracy of such images and reduces the segmentation error rate. At the same time, as the water surface image is usually affected by light intensity, water ripple and other factors, this paper also adopts the relevant correction algorithm to further improve the segmentation accuracy. [ABSTRACT FROM AUTHOR]

• A method for apple growth morphology based on Mask RCNN + PointRend has been proposed. • SPPF structure, CA attention mechanism and PANet are introduced as improvements. • CWD algorithm is used for model distillation to improve the detection speed. Identifying apples with different growth forms is an urgent task to be solved by harvesting robots on the path to intelligence. A method based on Mask RCNN + PointRend was proposed for identifying apples with different growth forms in the orchard in the work. The SPPF structure and CA attention mechanism were introduced to the backbone network of the Mask RCNN + PointRend algorithm to improve network performance because of the complexity of the apple orchard. Meanwhile, a bottom-up pathway was introduced into the feature extraction network to enhance feature dissemination, reuse, and fitting. Finally, the number of model parameters was reduced to improve the detection speed, and the CWD algorithm was used for the knowledge distillation of the model. The improved algorithms were called Mask RCNN-Point and Mask RCNN-PointKD. The recognition accuracy of the two enhanced algorithms on the test set was 85.65 and 74.13%, respectively. Besides, average detection time per image was 93.67 and 62.1 ms, respectively. The work provides a reference for identifying apples with different growth forms. [ABSTRACT FROM AUTHOR]

The recognition of apple fruits in plastic bags is easy to be affected by reflected and refracted light. In order to weaken the influence of light, a method based on block classification is proposed. The method adopts watershed algorithm to segment original images into irregular blocks based on edge detection results of R-G grayscale images firstly. Compared with the watershed algorithm based on gradient images, the segmentation method can preserve fruits edge and reduce the number of blocks by 20.31%, because graying image method, R-G, filters most of leaves and edge detection operator insures that the edge of fruits are detected accurately. Next, these blocks are classified into fruit blocks and non-fruit blocks by support vector machine on the basis of the color and texture features extracted from blocks. Compared with the image recognition method based on pixel classification, the proposed method can restrain the interference of light caused by plastic bags effectively. The false negative rate (FNR) and false positive rate (FPR) of the method based on pixel classification are 21.71 and 14.53% respectively. The FNR and FPR of the proposed method are 4.65 and 3.50% respectively. [ABSTRACT FROM AUTHOR]

Regarding the identification and location of bayberries in the natural environment, the work applied the dilated convolution to the res4b module of Mask RCNN backbone network—ResNet. The method was used to realize the accurate identification and segmentation of waxberry. First, we pre-trained the d -MRCNN network transformed by the dilated convolution with the COCO dataset. Then, through the migration learning method, the representative waxberry dataset was used to train the network for the identification and segmentation of waxberry. Finally, based on the same verification sample set, the work compared the Ostu and K-Means with the deep learning segmentation networks U-net and FCN. The result showed that the algorithm in this work was optimal, with the average detection accuracy and recall rate reaching 97 % and 91 %, respectively. It has high generalization in non-structural environment and better robustness with various forms. [ABSTRACT FROM AUTHOR]

• Apple image correction based on the adaptive Gamma method. • Using the basic image processing method to extract the complete fruit regions. • Obtaining the near-large fruit by the optimized iterative open operation method. It is appropriate for near-large fruit to be the first target of single-arm harvesting robot based on machine vision in each operation. The work introduced a method to obtain the near-large fruit from apple image in orchard. Firstly, the R-channel and G-channel images of orchard apple RGB (Red, Green, Blue) images were operated by the adaptive Gamma correction method to overcome the influence of natural illumination on subsequent image segmentation. After combination, the treatment process was formed by the basic image processing method such as corrosion, expansion, hole filling, and fixed threshold segmentation to obtain a complete, clean fruit area in the collected apple image. Finally, the near-large fruit was obtained from apple image of orchard based on the optimized iterative open operation. To verify the designed method, 30 apple images were selected according to strong to low light conditions from the captured images. Experimental results showed that the apple image after the adaptive Gamma correction was segmented to obtain the cleaner fruit region. The proposed method was used to obtain the correct near-large fruit with nearly real area. The optimized iterative open operation reduced the execution time of entire algorithm, with average and maximum reduction ratios of 70 and 84%, respectively. [ABSTRACT FROM AUTHOR]