Object Distance Measurement Using a Single Camera for Robotic Applications

Abstract

Visual servoing is defined as controlling robots by extracting data obtained from
\nthe vision system, such as the distance of an object with respect to a reference frame, or the length and width of the object. There are three image-based object distance
\nmeasurement techniques: i) using two cameras, i.e., stereovision; ii) using a single
\ncamera, i.e., monovision; and iii) time-of-flight camera.
\nThe stereovision method uses two cameras to find the object’s depth and is highly
\naccurate. However, it is costly compared to the monovision technique due to the higher
\ncomputational burden and the cost of two cameras (rather than one) and related
\naccessories. In addition, in stereovision, a larger number of images of the object need to
\nbe processed in real-time, and by increasing the distance of the object from cameras, the
\nmeasurement accuracy decreases. In the time-of-flight distance measurement technique,
\ndistance information is obtained by measuring the total time for the light to transmit to
\nand reflect from the object. The shortcoming of this technique is that it is difficult to
\nseparate the incoming signal, since it depends on many parameters such as the intensity
\nof the reflected light, the intensity of the background light, and the dynamic range of the
\nsensor. However, for applications such as rescue robot or object manipulation by a robot
\nin a home and office environment, the high accuracy distance measurement provided by
\nstereovision is not required. Instead, the monovision approach is attractive for some
\napplications due to: i) lower cost and lower computational burden; and ii) lower
\ncomplexity due to the use of only one camera.
\nUsing a single camera for distance measurement, object detection and feature
\nextraction (i.e., finding the length and width of an object) is not yet well researched and there are very few published works on the topic in the literature. Therefore, using this
\ntechnique for real-world robotics applications requires more research and improvements.
\nThis thesis mainly focuses on the development of object distance measurement
\nand feature extraction algorithms using a single fixed camera and a single camera with
\nvariable pitch angle based on image processing techniques. As a result, two different
\nimproved and modified object distance measurement algorithms were proposed for cases
\nwhere a camera is fixed at a given angle in the vertical plane and when it is rotating in a
\nvertical plane. In the proposed algorithms, as a first step, the object distance and
\ndimension such as length and width were obtained using existing image processing
\ntechniques. Since the results were not accurate due to lens distortion, noise, variable light
\nintensity and other uncertainties such as deviation of the position of the object from the
\noptical axes of camera, in the second step, the distance and dimension of the object
\nobtained from existing techniques were modified in the X- and Y-directions and for the
\norientation of the object about the Z-axis in the object plane by using experimental data
\nand identification techniques such as the least square method.
\nExtensive experimental results confirmed that the accuracy increased for
\nmeasured distance from 9.4 mm to 2.95 mm, for length from 11.6 mm to 2.2 mm, and for
\nwidth from 18.6 mm to 10.8 mm. In addition, the proposed algorithm is significantly
\nimproved with proposed corrections compared to existing methods. Furthermore, the
\nimproved distance measurement method is computationally efficient and can be used for
\nreal-time robotic application tasks such as pick and place and object manipulation in a
\nhome or office environment.