A Comparative Study of Structured Light and Laser Range Finding Devices

Transcription

1 A Comparative Study of Structured Light and Laser Range Finding Devices Todd Bernhard Anuraag Chintalapally Daniel Zukowski Abstract This is a survey of the resolution of data gathered by different 3D-imaging devices. The sensors covered in this paper consist of the Asus Xtion Pro Live [1], Microsoft Kinect [2], Microsoft Kinect with Nyko Zoom Lens [3], and Hokuyo URG- 04LX Laser Range Finder [4]. The results of this survey suggest that the Asus and Kinect (both with and without the Zoom Lens) possess similar capabilities, and are well-suited to near real-time applications, while the Hokuyo URG-04LX Laser Range Finder is able to detect smaller features than both the Kinect and the Asus, but is only viable for non-real-time applications. I. INTRODUCTION This survey was conducted with the goal of deciding upon the 3D-imaging platform for the Autoponics project [5] being conducted at Solid State Depot (SSD) in Boulder, Colorado [6]. Autoponics is a method of growing plants in which robots and other computerized machinery perform all aspects of plant food production: seeding, growing, harvesting, and processing. The Autoponics project at SSD is using the Robot Operating System (ROS) [7] framework for reading data and controlling system operations. Therefore, the ideal 3D-imaging device should have two important properties: it must be supported by ROS, and it must have a sufficiently high resolution to gather useful and reliable data. Specifically, the point cloud data produced by the device must be of sufficient resolution such that it can be used to produce a 3D model used for path planning and end-effector manipulation (i.e. harvesting parts of a plant). All of the devices tested in this survey are supported by ROS, and thus the ability to distinguish plants and their component features becomes the most important factor. The sensor package (which includes the imaging device) of the Autoponics system is mounted on a vertically-oriented cartesian robot, and so we will refer to a coordinate system throughout this paper with respect to the orientation of this vertical cartestian robot. Let us define the coordinate system such that the sensor is at the origin, pointed directly towards the positive Z-axis, which is parallel to the ground and orthogonal to the vertical XY-plane. The depth data is translated into Z-coordinates, and the resulting point clouds generated by the sensors create the basis for a varying surface that extends in the X- and Y-directions. Although the experiment is conducted in a similar manner for every device, it is important to note the differences in the purpose of each image sensor, as it may explain some of the results. The Microsoft Kinect was originally created as a enhanced gaming device for the Xbox platform. The primary purpose of the Kinect is to track players and their gestures. Since it is a mass market device, the cost of the sensor is relatively cheap at approximately $150. Another mass market device is the Asus Xtion Pro Live which costs about $190. The Xtion Pro was built with the purpose of sensing human movement for the PC platform. Since they are built to enhance an interactive experience, both the Kinect and Xtion Pro are real-time systems. The third image sensor used in the experiment is the Hokuyo URG-04LX Laser Range Finder. The URG-04LX has significant differences in function and purpose from the previous systems, which helps explain its price tag of almost $1,200. The URG-04LX is intended for use in indoor robot navigation and obstacle avoidance. It delivers a line of depth data in real-time, but creating 3D point clouds from this data requires post-processing, described below, and introduces enough latency that it is not considered a real-time system. The remainder of this paper is as follows: Materials, Methods, Results, Discussion, and Conclusion. II. MATERIALS The experiment conducted in this survey consists of the pieces seen in Figure 1. The three main components of the experiment are the 3D imaging devices, visual target, and the white cardboard backdrop behind the target. All three are placed on a flat eight-foot table. The visual target (Figure 2) is a one-foot by one-foot square of 1 4 inch-thick acrylic, laminated with an opaque brown material so as to eliminate the factor of transparency from the experimental variables. The features of the target are laser-cut holes that increase in size from right to left and from bottom to top. The diameter of the holes are 0.25 cm, 0.50 cm, 0.75 cm, 1 cm, 2 cm, 3 cm, 4 cm, 6 cm, 8 cm, and 10 cm. The Xtion Pro and the Kinect were tested without modification. Both of these devices use structured light to cast a grid of infrared points onto the environment. Because the light is distributed in two dimensions, each frame of the sensor data produces a complete point cloud. The Hokuyo Laser Range Finder differs in that it creates a 1D line of points at different depths, or a slice of a point cloud. To generate a full 3D point

2 topics, inserts the X-coordinates, buffers the modified laser scans, and then displays a collection of slices simultaneously in the ROS Visualizer (RVIZ) [9] to create the full 3D point cloud. III. METHODS We conducted two different tests in this survey the size of the smallest detectable feature and the depth resolution at increasing distances between the sensor and the target. The depth resolution demonstrates the minimum consistently detectable difference in depth of different features. Fig. 1: The test setup from the perspective of a sensor. In the foreground is the Microsoft Kinect with the Nyko Zoom Lens. Beyond it is the visual target and the backdrop. A. Smallest Feature Test The sensor was placed at one end of the table and the white backdrop at the other. We gradually moved the target from 10 cm from the sensor to 250 cm, taking screenshots and noting the distance each time a feature (the holes) was no longer consistently visible. Figure 4 shows a screenshot from this test. B. Depth Resolution Test In this test, the sensor was again stationary at one end of the table. The target is placed at increasing distances from the sensor, from 50 cm to 210 cm at 10 cm increments. At each distance, we placed the backdrop immediately behind the target and slowly increased the distance between the backdrop and the visual target until we could visibly differentiate between the backdrop and the features of the target. The depth data was mapped to the color spectrum, rather than simply gray-scale, making the difference more noticeable. A. Smallest Feature Test IV. RESULTS As distance increased, all devices exhibited an increase in the size of the smallest detectable feature, as shown in Figure Fig. 2: Diagram of the visual target. The outline is a 1ft by 1ft square. cloud from this data, we orient the laser such that it produces slices parrallel to the YZ-plane and take a series of slices by incrementing the X-position of the laser. The full point cloud is produced by programming a ROS node that stitches the slices together by inserting an X-coordinate into each YZ laser scan. The laser is mounted to a carriage on a Makerslide [8] rail, as shown in Figure 3. A stepper motor, also mounted to the carriage, controls a belt drive to move the carriage along the rail, and continually publishes its position to a ROS message topic. The stitching node reads both the position and laser scan Fig. 3: Picture of the X-Carriage on the Makerslide. The Hokuyo laser scanner (right, black and square) is mounted next to the stepper motor (center, black). Also visible is the IPEVO Document Camera (left, long and silver), which was unused in these tests.

3 Device Angular resolution (deg) Kinect Xtion Pro Live URG-04LX TABLE I: The published specifications of each device Device Linear Res at 1 m Smallest Feature at 1 m Kinect 0.16 cm 3 cm Xtion Pro Live 0.16 cm 2 cm URG-04LX 0.63 cm 1 cm TABLE II: The published specifications of each device distance, which we will call linear resolution, is given by L = D tan A Fig. 4: An example screenshot from the smallest feature test, using the Asus Xtion Pro at 60 cm from the target. Fig. 5: Size of smallest detectable feature at varying distances. Smaller values indicate better performance. 5. The Kinect performed worse than the other devices at every distance. The Kinect with the Nyko Zoom lens and Asus performed nearly identically, and the Hokuyo performed better than the other devices. However, the data for the Hokuyo is restricted to distances under 1.5 m. The three structured light sensors (Kinect with and without the Zoom lens and the Xtion Pro Live) all exhibit nearly linear growth until approximately 1.6 meters, at which point their performance degrades sharply. B. Ratio of Smallest Feature Size to Published Linear Resolution The minimum size of detectable features at given distances is closely related to the angular resolution of the sensors. The distance between two pixels of the point cloud at a given where L is the linear resolution, A is the angular resolution of the sensor, and D is the distance between the sensor and the target. While our experiments aimed to determine the best device for our application, we were also interested in comparing our experimental results with the published specifications for the devices we tested. According to the published specifications, we determined the angular resolution of each device, except the Nyko Zoom Lens 1, listed in Table I. As expected, the scattered light sensors do not generate any useful data at ranges less than roughly 0.5 meters. The Kinect claims to have a minimum effective distance of 0.8 m, however we were able to detect a 2 cm feature at a distance of 0.6 m. Similarly, the Xtion Pro Live claims a minimum effective distance of 0.8 m, and we detected a 1 cm feature at a distance of 0.6 m. Based on the angular resolution of the data, the linear resolution at 1 m for each device is presented in Table II, along with the smallest detectable feature size at 1 m. As a means to approximate the minimum number of pixels required to discern a feature, we used the ratio of the experimentally-derived minimum detectable feature size over the theoretical linear resolution calculated from published device specifications. Figure 6 depicts this ratio as a function of distance. To detect a feature, the Hokuyo required the feature to be 2-5 times larger than the linear resolution at a given distance, whereas the Kinect and Asus required features to be approximately times larger than the linear resolution at a given distance. C. Depth Resolution Test Of the three devices tested, the Asus exhibited the best depth resolution, while the Kinect with Nyko Zoom exhibited the worst depth resolution. Figure 7 charts the depth resolution over distance for all three devices, and includes a linear regression fit for each data series. 1 Nyko has not published any specifications regarding the angular resolution or field of view of the Kinect with the lens attatched.

4 V. DISCUSSION Fig. 6: This graph displays the ratios of the sizes of the smallest detectable feature at a given distance over the theoretical linear resolution at that distance. This quantity is an approximation of the number of pixels necessary to distinguish a feature from its surroundings. Fig. 7: The depth resolution test yielded highly variable results, but least-squares regression shows that there are consistent trends within the data. The results of our experiment suggest that the Hokuyo URG-04LX Laser Range Finder has superior feature-detection ability when compared to the Xtion Pro, Kinect, and Kinect with Nyko Zoom Lens. Surprisingly, there was little difference in performance between the Kinect and Kinect with Nyko Zoom lens structured light sensors. The Xtion Pro Live outperformed the Kinect, both with and without the Zoom lens, in both tests, particularly in the depth resolution. The Nyko Zoom proved to be a double-edged sword, improving the Kinect s smallest feature detection to the level of the Xtion Pro Live, but negatively affected its depth resolution. All of the image sensors had a significantly lower usable resolution than what was listed. Although this is to be expected based on the observations that it takes more than one or two pixels to distinguish a feature in a point cloud, and the resolution listed by the manufacturers is just the real-world distance between each observed point. In practical applications, the minimum detectable feature size for the Kinect and Asus is times larger than the linear resolution at a given distance. In this regard, the Hokuyo performed closer to published specs, requiring minimum detectable feature sizes of only 2-5x what the published specs suggested. Users of the Kinect and Asus devices should consider this large discrepancy when choosing to use either sensor for feature recognition applications. However, the poorer resolution does allow for a faster sampling rate and greatly reduced cost. For the purposes of the Autoponics project, realtime sampling is not critical, and accurate detection of small features is important; thus, the Hokuyo Laser Range Finder is a more appropriate device for mapping and feature detection in our application. There were a number of possible sources of error in the experiments. The measurements included some degree of subjectivity, as each data point required a decision by the experimenter. During the smallest feature test, the holes frequently flickered within a small range of distance, rather than blinking out immediately. To maintain consistency, we waited until the end of this range, though in some cases it was an asymptotic falloff instead of a clear end. In addition, at longer distances the smaller features shrank to only a few pixels in diameter, that may have been detectable to a finely tuned algorithm, but not to the human eye. In future testing, we intend to use an image processing algorithm rather than a human operator to create static thresholds and will conduct more trials for more reliable averages. Similarly, the method used to test the depth resolution required human judgement. The depth data was mapped to the color spectrum, and thus the test relied on the experimenter to decide when the backdrop seen through the holes of the target was a different hue than the target itself. This proved to be a highly subjective metric, as evidenced by the high variance of our results. We suspect the error may have varied with the color spectrum (i.e. the differences in hues of green were more difficult to distinguish than hues of blue); by depending on color, slight color blindness in the experimenters may have

5 contributed minor error as well. Again, in future testing we will use an image processing algorithm for more consistent results, in addition to more trials. By varying the offset of the color spectrum during the trials such as cycling from green to green, then red to red, then blue to blue we can eliminate the bias due to particular hues. The experimental setup may have also contributed minor errors. It is difficult to align the visual target at the correct angle to the image sensor. Both the laser from the URG-04LX and the field of view for the Kinect and Xtion Pro must be perfectly orthogonal to the target acrylic board to get accurate measurements. Although we used a level to orient the target acrylic and the image sensor, our equipment was not accurate enough to ensure perfect angles for any measurement. Though the offset was minor, it likely affected our results. Finally, the visual target was composed of features of constant size, and in the smallest feature test this limited the number of possible measurements. To remove this barrier, one could construct a target with a window of variable size. This device would allow for continuous measurements as one increases the distance between the sensor and the target. REFERENCES [1] Asus xtion pro live, Sensor/ Xtion PRO LIVE/. [2] Microsoft xbox kinect, [3] Microsoft xbox kinect, [4] Hokuyo 04lx-ug01 laser scanner, 07scanner/urg 04lx ug01.html. [5] The autoponics project, [6] Solid state depot, [7] Ros (robot operating system, [8] Makerslide, [9] Rviz, VI. CONCLUSION In this paper, we have presented and discussed the materials, methods, and results for surveying different sensors used to generating point clouds for mapping and feature detection. We conducted what we believe to be a fair experiment in real-world conditions, using four different systems that are available on the market today. Published sensor specifications do not necessarily provide enough information to evaluate the efficacy of a device for a specific application. While the linear resolution of the Kinect and Asus look impressive on paper, the practical use of these devices for feature detection falls short of what one might expect from the published figures. The data we collected provides a perspective on the feature detection abilities and depth resolution that can be expected when using these devices in a real-world environment. From this data, we were able to determine that, among the devices tested, the Hokuyo URG-04LX Laser Range Finder is best suited for use as the primary tool for feature detection and localization for the Autponics project. In addition to superior feature detection abilities, the URG-04LX is able to provide useful data at short distances, unlike the structured light sensors. Although our experiment was inherently subjective in the techniques used to collect data, the URG-04LX was clearly superior in terms of its ability to recognize small features at near and far distances. ACKNOWLEDGMENT For acting as a mentor, and providing valuable guidance to the Autoponics project team, we would like to thank Dr. Correll from the. We also want to acknowledge members of the Solid State Depot for providing an unique perspective on both our experiment and the Autoponics system.