This PDF file contains the front matter associated with SPIE Proceedings Volume 12452, including the Title Page, Copyright information, Table of Contents and Conference Committe list

Garbage pollution is a very difficult problem in environmental governance. Due to the many sources of garbage pollution and a wide range of impacts, this problem is only slow to solve by human means. In order to improve the automation of garbage disposal, on the one hand, this paper proposes a garbage detection method based on CNN (convolutional neural network) using multi-layer feature processing. On the other hand, the detection algorithm is combined with an industrial robot to form a complete garbage sorting system. This paper uses the one-stage idea to first optimize the backbone structure to improve the extraction effect of shallow features. Then the attention module is introduced to make the network pay more attention to information that plays a key role in garbage detection. Finally, a multi-layer feature fusion method is used to combine the features of the shallow network with the features of the deep network to generate a fused feature map for use in target detection tasks. The experimental results show that the detection speed of the method proposed in this paper is 13.75% higher than that of SSD, and the garbage detection accuracy reaches 99.5%, which is better than the SSD detection algorithm. The garbage detection method proposed in this paper can quickly realize the precise positioning of garbage and complete automatic robot sorting.

In this paper, an image saliency detection method based on regional label fusion is proposed to solve the problems with fuzzy boundaries, unclear profile, and less interior density commonly existing in the researches of salient region detection. The image is segmented by super pixel segmentation algorithm, then the spectral clustering is carried out for the super pixel region to reduce the number of regions, thereby the label set with the boundary information could be obtained. Next, three salient features of the image have been fused under conditional random field model to generate the coarse saliency map. Afterwards, regional label fusion method is operated, which organically fuses the boundary information into the coarse saliency map by using the salient mean value calculated with the label information as the regional salient features, moreover, together with adaptive threshold segmentation algorithm to acquire reconstructed saliency map. At last, accurate salient region detection is achieved by calculating with a tag indicating vector defined and reconstructed coarse saliency map. Experimental results show that the salient regions obtained by this algorithm display clearer boundary contours and that the density of salient regions has been greatly improved compared with the other six significant detection methods prevailed in recent years.

The traditional rapid expansion random tree (RRT) algorithm has poor efficiency in the motion planning of the manipulator. Based on the traditional RRT algorithm, this paper introduces the target offset strategy in the process of expanding leaf nodes. When the algorithm falls into a local minimum, it will select the expansion point, so as to quickly break away from the minimum. The improved RRT algorithm and other algorithms are simulated in Mathematica. The experimental results show that the improved algorithm can guide the growth direction of the tree, improve the convergence speed of the algorithm, make it difficult to fall into local minimum, and improve the motion planning efficiency of the manipulator in simulation.

In order to detect the object and inspect the road conditions in real-time, the 2-dimensional (2D) and 3- dimensional (3D) data obtained from the onboard sensors, LiDAR and digital cameras are analyzed for object recognition to assist driving. Due to the uncertainties of the dynamic objects, such as pedestrians, animals or vibrated vehicles, extraction of complete and clear objects from LiDARs datasets requires complex post-processing since LiDAR data can be used for scanning at long distances, i.e., 300m, which can alarm the driver timely to take necessary actions. The dynamic and static objects from the LiDARs point clouds can be detected with the teacher-student framework algorithm along with the KITTI dataset. Furthermore, a semi-supervised theory is utilized to improve detection performance.

In multi-robot systems with dynamic and complex environments, robots are required to avoid not only the static objects but also other moving robots. To solve this problem, this paper presents an implementation of cooperative collision avoidance architecture based on optimized sampling-based collision avoidance paradigm. In our work, localization error is considered and bounded in adaptive Monte-Carlo localization process. Plus, we employ velocity obstacle paradigm in predicting collisions. Subsequently, by using Sampling-based planner and optimization theory, we get an optimizing velocity selection policy. Furthermore, we also introduce our distributed multi-robot system model in this paper. By applying the cooperative collision avoidance method in Gazebo self-driving car simulation environment and ROS mobile robots, it is illustrated that our approach is applicable and well-performed.

In this paper, a swarm spherical robot based on machine vision is proposed to adapt to the special applications of robots in the fields of land-based safety, navigation, surveillance and exploration, and discusses the spherical robot in terms of mechanism, control principles and positioning, respectively. The spherical robot structure is a telescopic low-degree-of-freedom parallel metamorphic mechanism, which is first analyzed mechanically using a modified Grubler-Kutzbach criterion to determine the degrees of freedom and to analyse the mechanics of key components. Secondly, the swarm robot consists of multiple spherical robots connected by magnetic forces. By designing a relevant path planning scheme to control the sequence of movements of these spherical robots, the movement of the whole swarm in the direction of a specific point can be achieved. Finally, the positioning and wireless control scheme of swarm robot based on machine vision is presented, and the motion principle of swarm robot is briefly introduced.

In this paper, the ultrasonic transmission process and damage detection in rod structure are experimentally studied. In the finite element simulation, according to the nonlinear interaction between the different Lamb wave excitation signals and the fixed size crack on the rod, the signal mechanism is analyzed and how to locate the damage signal is explained. The fatigue crack evolution of rod assembly is further tested by experimental method. It is consistent with the simulation results. It is found that the ultrasonic guided wave signal with excitation is very sensitive to cracks. Finite element analysis and excitation signals can be used as potential methods for quantitative evaluation of cracks in metal parts. The results of this study provide a basis for modeling and simulation using finite element in the field of nondestructive evaluation and structural health monitoring

The main loads and forces on the tire are carried by its cord-rubber composites structure. The geometry of the cord rubber composites needs to be measured during the production of the tire to ensure its quality. In this paper, a vision-based high-speed cord-rubber composites measurement system is developed. The system can accomplish the high-precision calibration of the linear camera based on the Checkerboard calibration board. Integration of the infrared light source into the system to improve image quality at high-speed motion. Based on this, the edge extraction and geometric parameter calculation of the measurement object are realized by using the Hough transform and GPU acceleration algorithm. Finally, the system is verified by measuring a standard sample and comparing the measurement results with the standard values. The standard deviations of the two angle measurements are 0.004° and 0.014°.

In recent years, the rapid development of deep learning makes it more and more widely used in the field of defect detection. Compared with the traditional machine vision methods, the deep learning methods based on Convolutional Neural Networks (CNN) have stronger feature learning abilities and can achieve higher detection accuracy and work efficiency in the field of surface defect detection of industrial products. However, supervised deep learning algorithms require a large amount of labeled data, making it difficult to generalize practically. To this end, we propose an unsupervised defect detection method MSFR-VAE for Multi-Scale Feature Reconstruction-Variational Auto Encoder: It realizes defect detection and localization by reconstructing the deep features of the input image and only needs to be trained on normal samples. Different from the image-based reconstruction, the feature-based reconstruction method can make the model focus more on the key features that can distinguish the normal and defective samples, so as to improve the detection effect. Besides, we use the pre-trained CNN for Multi-Scale feature extraction which is carried out from an image pyramid to detect defects of different sizes. Moreover, in order to make full use of the deep features, we use Variational AutoEncoder (VAE) to learn the feature distribution of normal samples for better reconstruction. Extensive experiments on the challenging and newly proposed MVTec AD dataset show that our method outperforms baselines.