China
Africa
Explore chapters and articles related to this topic
Sensor Fusion in Multiscale Inspection Systems
Published in Wolfgang Osten, Optical Inspection of Microsystems, 2019
All robust fitting algorithms can be achieved by applying Random sample consensus (RANSAC)-based methods in order to remove outliers. The final estimation of the line or plane coefficients is obtained by a least-square fit.
Saliency and spatial information-based landmark selection for mobile robot navigation in natural environments
Published in Advanced Robotics, 2019
Gábor Kovács, Yasuharu Kunii, Takao Maeda, Hideki Hashimoto
A robust scale-space method is used to extract multiple feature points for visual odometry. Based on our experiments regarding the ability to provide an adequate number of feature points, we are using CenSurE (Center Surround Extremas for Realtime Feature Detection and Matching) as a feature detector and FREAK as a descriptor [16,17]. FREAK is a binary descriptor that provides fast feature point selection during real-time operation. Feature points are extracted from 2 to 10 [m] range where the acquired 3D data is more accurate. RANSAC (RANdom SAmple Consensus) [18] algorithm is used to remove outlier data from the feature points as a preprocessing step to reduce measurement errors. The movement compared to the previous image frame is calculated by geometric transformations in the Cartesian coordinate system for all six degrees of freedom.
Activity representation by SURF-based templates
Published in Computer Methods in Biomechanics and Biomedical Engineering: Imaging & Visualization, 2018
Md. Atiqur Rahman Ahad, JK Tan, H Kim, S Ishikawa
The Motion History Image (MHI) is a very well-known method for action recognition and understanding. Though the MHI method is simple to compute in representing video actions into a single image, it has some constraints that hinder it for challenging applications. One major concern of this method is its inability to resolve the motion self-occlusion problem, which appears due to motion overwriting (Pantic et al. 2005; Meng et al. 2006, 2007; Ahad et al. 2009a, 2009b). Also, its performance is reasonable when the action is simple and having a single person in the view. Therefore, we ponder to solve this issue to overcome this problem so that the complex actions and two-person interactions can be recognised better. In this paper, we explore this area. Here, we propose a spatio-temporal method so that it can recognise different complex and usual activities. We compute locally produced interest points and hence, calculate global features for action representations. For this purpose, Speeded-Up Robust Features (SURF) is extracted as key interest points. The SURF is a scale-invariant as well as rotation-invariant interest point detector and descriptor (Bay et al. 2008; Ehsan & McDonald-Maier 2009; Schweiger et al. 2009). On the extracted key points, we employ optical flow. To reduce outliers, we apply RANdom Sample Consensus (RANSAC). The gradient-based optical flow vectors are separated into the four separate motion vector channels called, (i) up-direction, (ii) down-direction, (iii) left-direction and (iv) right-direction. We compute history images and the corresponding energy images from these optical flow channels. In the final stage, a frame-subtracted accumulated image is also considered, which is masked in order to remove unwanted corner points in the scene, especially in the case of outdoor cluttered scene. Through these steps, we obtain the SURF-based History Image () and the corresponding Energy Image, which can demonstrate a better image representation. Hu moment invariants (Hu 1962) are computed from these image templates for each action to create feature vectors. We exploit a nearest neighbour classifier and leave-one-out cross-validation as partitioning scheme. In this paper, we show comparative results of our method with similar other approaches and demonstrate that the proposed method shows satisfactory recognition results.