Object recognition technology for autonomous unmanned control based on stereo data analysis

Abstract:

The paper represents a description of the object recognition technology on a road using moving stereo pair of cameras in real time. The developed algorithm is characterized by high processing speed and performs reliable detection of arbitrary shape obstacles. Calculation time for one iteration from receiving input images (size 1280 ∗ 480) to detecting obstacles is less than 2 msec for a Geforce GTX 1080T video card.

Introduction

Object recognition is an important step in solving problems of identifying obstacles by unmanned vehicle control and movement of robots, in object detection at production processes automation. One of the main parts of the object recognition system is stereo vision, which reconstructs a three-dimensional scene of the visible area with real metric values by a set of images. The restoration of a three-dimensional image is used to solve a plenty of applied problems in various activities [1, 2]. In the motor-car industry stereo vision is used to recognize obstacles and analyze the quality of paving. Approbation of the unmanned vehicle projects [3] determined limitations of object recognition algorithms such as movement in urban conditions, obligatory presence of road markings, recognition of a limited number of object classes require revision of the proposed solutions. A separate direction is the detection of moving objects, which accuracy is based on the analysis of programs working with open initial code does not exceed 60% [4]. Using stereo vision allows to solve two important tasks for unmanned control – object detection and obtaining an estimate of the distance to them, as well as determining their size. A detailed review of the stereo vision algorithms based on the analysis of two images is represented in paper [5]. The resource capacity of the algorithms does not allow them to be used to calculate stereo data in real time. The method proposed in paper [6] is based on the control points searching, which reduces the amount of data processed. However, the described algorithm implementation by CPU does not allow it to be used to detect objects in real time when using input images with high resolution.

Читать статью