Search

A new dataset has been assembled to compare the efficiency of simultaneous localization and mapping algorithms out of doors in areas with high a priori uncertainty. The images were made using a high-resolution stereo camera, with reference camera trajectories being used to calculate movement within the scene based on manually identified key points. The resulting dataset of images can be used to compare the robustness of simultaneous localization and mapping algorithms under complex lighting conditions outdoors.

A method for object recognition based on the use of a structural description of images of three-dimensional scenes is proposed, and it is experimentally compared with existing analogues. A quantitative assessment of the recognition accuracy and speed of the studied methods is presented. Methods for detecting key points and building image descriptors in three-dimensional space are studied. The algorithm developed here can be used to address issues with mobile robot and unmanned aerial vehicle navigation under the conditions characterized by a high degree of a priori uncertainty of the scene.

This paper discusses the problem of automatically classifying visual stimuli (animate and inanimate objects filtered at high and low spatial frequencies) using an observer’s electroencephalogram. Classical machine-learning methods (a support-vector machine that employs, among other things, wavelet attributes) and convolutional and recurrent deep-learning neural networks were used for the classification. The recognition accuracy was analyzed as a function of the selected classification methods, the placement of the electrodes, the time intervals, and the problem to be solved. The results show that the classification accuracy is 79% for sharp/smeared images, 61% for animate/inanimate objects, and 50% for classifying four classes of images.

This paper discusses surface-segmentation methods in three-dimensional space in terms of a solution of the problem of structural comparison of images. An algorithm is presented for comparing images in three-dimensional space, and methods are considered for making the developed algorithm more stable against the picture-taking conditions. A quantitative estimate is given for the segmentation accuracy and the response rate of the methods being studied. Based of the results of the investigation, a modification of the image-comparison algorithm is proposed.

This paper analyzes the extent to which a semiglobal algorithm for stereo vision is stable against the picture-taking conditions and the characteristics of the object of observation in the soft-approach problem. It is shown that the algorithm under investigation is stable against various types of background under different illumination conditions and in the presence of noise. For long distances, it is recommended that the internal parameters of the algorithm undergo fine-tuning by means of machine-training methods. A technique is developed that makes it possible to establish the stability of the algorithm against shapes of the object of observation that vary in complexity.