Fuzzy decision machine (FDM) compared with finite state machine is proposed in this article to verify the appropriate performance on the soccer robot platform. Fuzzy logic with the flexible membership function adaptively produces the robust strategy for real-time reaching the better decision action for soccer robots working in the higher competition environment. This study feeds actual physical parameters into the fuzzy soccer robot system by a Gazebo simulator to ensure the perfect visibility. Therefore, the deployed 3D Gazebo scene for the specific Federation of International Robot-soccer Association (FIRA) RoboSot competition fields is greatly supplied with soccer robots as a friendly support. Moreover, four strategies are inspirited from experts’ experiences to implement the strong fuzzy rules through Robot Operating System (ROS). Fuzzy decision system infers the attacked or defended mode with the respective robot state information to select the near optimal robot action. The primary team’s object is to quickly attack the opponent’s goal when the fighting condition is favorable for the team. In the other case, the tactic is changed to reduce the probability of intercepting by the backswing way. FDM compares with the finite state machine in soccer robot competition fields to verify its availability. This experiment shows that the addressed FDM transforms the expert’s knowhow into appropriate fuzzy rules to improve the winning rates. Therefore, the generated soccer robot system contains the ability to make the best reactions in real time for approaching the desired home goals. The ROS-based system fuzzy adaptive machines embody the faculties to approach the performed strategy in real platform for getting the higher wining rate in the dynamic, complex and uncertain soccer robot competition games.

In the field of biometric recognition, convenience and security of the system are highly demanded. A large database usually leads to long response time and high computational complexity. This work, a new method, is presented to extract the vein patterns from near-infrared images, which are enhanced through the directional wavelet transform and the eight-directional neighborhood methods to further reduce the required computational cost as well as to preserve key information from low-resolution images. In addition, the region of interest of the finger-vein is also robustly located with the physiological properties of a human finger. As a result, a database composed of 340 images was employed to generate the required training and testing of vein geometrical features, that the proposed system can yield real-time requirement by achieving 0% false accept rate, 0.25% false reject rate, and recognition rate up to 100%. Meanwhile, the response time is 300 μs, and thus the proposed algorithm is a very effective candidate for a personal identification system.

This article proposes highly autonomous map generation and path navigation based on the Robot Operating System (ROS) platform. The mobile robot concurrently completes visualized map generation and path navigation even in an unknown environment. Autonomous visualization robot systems combine the Simultaneous Localization and Mapping (SLAM) and dynamic search techniques to self-drive to any desired target. The Hector SLAM is applied with only one LiDAR to continuously extract high-accuracy information from grid maps of neighboring environments. Due to the related robot radius, the grid maps are flexibly approximated by weighted scalar formulas. Then, the novel hybrid neighboring and global path planning is determined to achieve the appropriate position for fitting mobile robot navigation applications. In neighborhood search, the A* algorithm first explores the shortest path selection between robot and target with the perceptual information of the LiDAR. Global path selection with the dynamic window approach (DWA) is applied to improve the previous neighborhood search of the A* algorithm. The DWA accurately predicts all possible moving paths and chooses the best path planning. The mobile robot follows the shortest path and avoids obstacles to achieve the appropriate target. Based on repeated executions, the mobile robot explores its neighboring block and updates into global maps. The global path-planning scheme is restarted if the robot finds obstacles. This strategy allows robots to fit the appropriate maps, and to quickly react and effectively avoid the danger when they encounter some unexpected conditions. Several mobile robot navigation experiments illustrate that the autonomous path-planning and self-localization abilities can achieve the desired goals through the support of the flexible ROS platform. It is expedient to rebuild the visualized maps for the appropriate mobile robot applications even in unknown, unusual and complicated environments.

In this article, two key techniques, a multiple filter set and a feature classification method based on support vector machines (SVMs), are proposed for fingerprint feature identification. The Gabor filter (GF) has been verified as a successful method and is usually selected for fingerprint feature detection. Although the GF has high accuracy in feature extraction under a wide angle representation range, not all degree ranges of the filter are necessary. One disadvantage of different degree ranges is high computational costs. Moreover, fingerprints have clean direction field representations, which can be used to design suitable filter sets with low computational complexity. This article adopts modified Haar-like patterns to perform near-circle filter sets for acceptable feature coverage. In the processing of overlapping fingerprint automatic detection in a whole image, this article proposes an efficient method based on two statistical results, the mean and standard deviation in the frequency domain response under the discrete wavelet transform. To separate overlapping fingerprints, this article adopts the Gaussian matrix and discrete Fourier transform, where the correct angle is decided by automatic fingerprint detection. Through feature identification, this article proposes a multiple filter set for feature extraction and the efficient SVM-based classifier. In a performance comparison, using ten Haar-like patterns and a cascade classifier, which had a built-in open CV library as a benchmark, the proposed algorithm can reduce approximately 50% of computations on average and maintain an equal accuracy.

Margin is a significant algorithmic parameter that has an impact on the performance of margin-based machine learning algorithms. Margin setting algorithm (MSA) is a novel margin-based learning algorithm. However, there is no comprehensive study concerning the impact of setting margin on enhancing the performance of MSA. In this article, we studied the impact of margin on performances of MSA by comparing it to another popular margin-based algorithm, the support vector machine (SVM). This comparison comprehensively analyzes and compares how margin affects training performance and generalization, both theoretically and experimentally. In our theoretical analysis, margin definition and margin impacts are comprehensively discussed by demonstrating how they affect the decision boundary. Experimental analysis is performed on two-dimensional Gaussian data sets and benchmark data sets. The experimental results support our theoretical analysis, revealing that with an increased margin, training performance gets worse and generalization tends to improve within a certain range.

Illuminant estimation is a process of finding light source color chromaticity to achieve color constancy. A number of combinational color constancy algorithms that combine the estimates of multiple statistic- and learning-based unitary algorithms now appear in the literature. Since the unitary algorithms fail to estimate the illuminant for a wide range of scenes, the combination method aims to fuse the estimate of different unitary algorithms to achieve robust illuminant estimation. The traditional combinational method either uses a static weight to combine the estimate of unitary algorithms or choose a best unitary algorithm for the input image. The former one fails due to the use of static weight which is biased to the dominant scenes in training data and the second one has the difficulty to train a multi-class model with limited training data. This article addresses the limitation of combinational methods and proposes a novel multi-class dynamic weight model. The proposed method classifies images into different groups and a distinct dynamic weight generation model (DWM) is used by each group for the generation of weight to combine the estimate of unitary algorithms. The DWM proposed in this article generates the combination weight using an image feature and a static weight coefficient. Therefore, the weight changes with the image characteristics. Since the proposed method generates the combinational weight based on the image feature and it uses different DWMs for each image group, the proposed method is able to successfully combine the estimate of unitary algorithms for a wide range of scenes. Extensive experiments on Gehler–Shi and NUS-8 camera color constancy benchmark data set shows the proposed method is able to improve the performance by nearly 37% and 62%, respectively.