3.1. Temporal Intensity Variation
Five complete scans of the 75% reflectance spectralon panel are shown in Figure 2
, which indicates more than 5 hours of scan operation for a fully charged battery for each scan. The means of both the grey intensity [Figure 2(a)
] and the raw intensity [Figure 2(b)
] increases with increasing battery time. The intensity range, defined as the difference between the maximum and minimum mean intensity of one single complete scan, is less than 17.4 and 402.7 for the grey and the raw intensities, respectively. The standard deviation of the grey intensity [Figure 2(c)
] and the raw intensity [Figure 2(d)
] also increases with increasing battery time, and the variance of the grey intensity is significantly less than that of the raw intensity. When the standard deviation reaches its maximum (usually after 3 hours of battery time), the ratio of the standard deviation to the mean of the grey intensity is ∼0.06, and that of the raw intensity is ∼0.08. The smaller ratio value indicates a smaller intensity variation, which is more preferable for segmentation applications, and thus, the grey intensity should be used in this case.
Usually after approximately 1.5 hours of battery time, the ILRIS3D display screen indicates the battery is at “LOW” status. Considering only data collected when the battery is “LOW”, the intensity range of grey and raw intensities are reduced to less than 3.9 and 87.2, respectively. For the grey intensity, the variation is reduced to 22.4% (= 3.9/17.4 × 100%) of the original value, and this can be interpreted as a warm-up time increases the stability of ILRIS3D intensity.
The nominal warm-up time of ILRIS3D is 0.5 hours [17
]. However, our experiment results show better stability of the grey and the raw intensity after 1.5 hours [Figure 2
]. Unfortunately, this warm-up phenomenon cannot be modeled or predicted. A proprietary mechanism called ASC, Automated Scan Correction, is implemented by the manufacturer to compensate for temperature drift of range and angular measurements, but not for intensity [17
]. The access to the information regarding ASC is restricted by the manufacturer. We are unable to assess whether ASC is useful for intensity related use, i.e. predicting the warm-up phenomena of the intensity. Consequently, all the data collected afterwards are at battery “LOW” status to obtain consistent LIDAR intensity.
3.3. Planar Surface Scans
For the white surface experiment [red rectangle in Figure 1(a)
], the concentric circular pattern, is prominent in the intensity image sequence from 200 to 253, shown in Supplementary Material 1, with a repeating frequency of every 5 or 6 grey intensities. Intensity images of 206–217 and 255 are excerpted from the intensity image sequence for illustration and are shown in Figure 3
. For this experiment, except for the empty intensity image, such as 207 and 216 shown in Figure 3
, where on lidar point cloud is collected at this intensity value, there are approximately 4 concentric circles in each intensity image. The intensity images of 206–211 show a complete cycle of the evolving pattern for a planar surface obtained by ILRIS3D, where the radii of the circles are increasing with increasing grey intensity values, while the most inner circle starts (grey intensity of 206) as a fuzzy circle encompassing a large amount of lidar points. Another cycle of the evolving pattern can be seen with the intensity images from 212 to 217 in Figure 3
. The intensity image sequence in Supplementary Material 1 shows a total of 9 cycles of evolving pattern of the concentric circles for the white surface experiment, where the starting intensity images are 200, 206, 212, 218, 223, 229, 234, 241, and 247, respectively. Although the evolving pattern of the concentric circles is different in details for the 9 cycles (Supplementary Material 1), they are consistent in general appearance and can be easily identified. Given the grey intensity format (of 8 bit dynamic range) employed by ILRIS3D, the saturated signals are expected to be contained in the intensity image of 255, shown in Figure 3
, where lidar points are scattered all over the white planar surface, and doesn’t provide distinctive visual cues as those of the concentric patters in intensity images of 206–253. This result from the white surface experiment suggest that the visual cue of the varying radii can be used a criteria to segment ILRIS3D data into a planar surface.
This concentric circular pattern is due to the internal process of intensity scaling and conversion from 16 bits to 8 bits by the manufacturer design [17
]. The grey intensity images of 210–224, which represent a complete cycle of the evolving pattern, and the intensity image sequence of a concrete fence painted with two different colors [Figure 1(b)
] are shown in Figure 4(a)–(e)
and Supplementary Material 2, respectively. Because the fence is 5 m away from the ILRIS3D and the height of the wall is 1.8 m, only one-half of the concentric circular pattern with varying radii is visible for surface I; an even smaller portion of that is visible for surface II [Figure 4(a)–(e)
and Supplementary Material 2]. For surface I, the concentric circular pattern is half circles and the centers of these half circles appear to be at the top edge of surface I, which can be seen in grey intensity images of 210, 227, and 244 shown in Supplementary Material 2. For surface II, only an arc-like feature can be recognized, and the centers of these arcs appear to be located slightly below the top edge of surface I. The repeating frequency is 17 grey intensities, which is different than that of the white planar surface. And, only one-third of the intensity images contain the 3D laser points, i.e. except for grey intensity of 187, 190, 193, etc, all intensity images are empty. The radii of the concentric circles, shown in Figure 4(a)–(e)
and Supplementary Material 2, do not increase with increasing intensity values as the white surface experiment (Figure 3
). For example, the evolving pattern of the most inner circle of surface I, denoted by the red arrows in Figure 4(a)–(e)
, experiences the following stages within a complete cycle: starts as a small fuzzy circle [Figure 4(a)
], becomes a large circle [Figure 4(b)
], evolves to a band-like feature [Figure 4(c)
], turn back to a circle [Figure 4(d)
], becomes a band-like feature with large width, again [Figure 4(e)
]. The evolving pattern of the varying radii can be more easily identified with the use of intensity image sequence in Supplementary Material 2. The presence of ring- and arc-like features, change of repeating frequency, and the empty grey intensity images are all due to the internal process of intensity scaling and conversion from 16 bits to 8 bits [17
Following the results implied by the white planar surface experiment (Figure 3
), that the point clouds of the varying radii can be segmented into the same surface, we are able to confirm the existence of surfaces I and II by observing the intensity image sequence in Supplementary Material 2. A linear distinctive discontinuity feature can also be found at the boundary of the two surfaces [denoted as red rectangle in Figure 4(a)
] due to different color paintings of the two surfaces, which substantiates segmenting the point cloud into two planar surfaces. A GUI (graphic user interface) program is developed in Matlab (MathWorks, Inc.), which facilitates the manual selection process of segmenting the lidar points into two surfaces, and the result is shown in Figure 4(f)
, where red and yellow denote surfaces I and II, respectively, while the un-segmented 3D laser points are denoted as blue.
Because the grey intensity range for surfaces I and II are 187–251 and 190–251, respectively, a simple threshold is insufficient to separate the point cloud of the two surfaces, even though they are visually different in color.
The results from the scan of the four planar surfaces [Figure 1(c)
] are shown in Figure 5
. The grey intensity ranges for surfaces I–IV, shown in Table 1
, are smaller than those tested in previous experiments. Thus, the repeating frequency can not be determined for surfaces I–IV of this experiment. For surfaces I, III, and IV, the smallest height and width are 1 m and 1.5 m, respectively; the height and width of surface II is 0.5 m and 1.5 m, respectively. The size of each surface is smaller than those tested in previous experiments. And only a quarter or less of the circular pattern is visible [Figure 5(a)–(c)
and intensity images sequence in Supplementary Material 3]. For this experiment, the boundaries separating different surfaces, denoted as red rectangles in Figure 5(a)–(c)
, can be distinctively identified, which is helpful in lidar point segmentation.
For surfaces I, III, and IV, the circular pattern is large enough to be recognized as different surfaces. Due to the small dimensions of surface II, the concentric circular pattern is not prominent. The identification of surface II can be facilitated by the presence of the boundaries, which are linear features and denoted as the red rectangles in Figure 5(a)
. The red rectangle shown in the middle of Figure 5(a)
is a boundary separates two groups of varying radii which represent surfaces III and II, respectively. Due to limited number of grey intensity, the concentric circular pattern of surfaces I and II do not appear in the same intensity range (Table 1
), and the other red rectangle shown in Figure 5(a)
is the top boundary of surface II, where the same boundary (which becomes the bottom boundary of surface I) can be found as the bottom-right red rectangle in Figure 5(c)
. The other boundaries can also be found for surface IV in Figure 5(b)
, and surface I in Figure 5(c)
. Thus, in addition to the visual cues of the concentric circular pattern, the linear boundaries, which separates different group of varying radii or delineates the boundary of a group of vary radii, are helpful for the segmentation task.
shows the segmentation results using the visual cues of the concentric circular patterns of these small surfaces and boundaries, appearing as linear features dividing each concentric circular pattern shown in the intensity image sequence in Supplementary Material 3, using the GUI program by manual selection. The order of which grey intensity value is processed is of no importance because they will be distinctively different groups of lidar points of varying radii in the intensity image sequence.
The information of displacement and variation in the point cloud data is usually employed by geometric based segmentation algorithm [2
]. Figure 5(e)
shows the nadir view of a subset of point cloud extracted from the black rectangle in Figure 5(d)
, and the red arrow indicates the boundary between surfaces I and III. It is evident in Figure 5(e)
that the variation and displace of point clouds for the two surfaces are similar, which implies a geometric based segmentation algorithm is expected to produce non-satisfactory results because the information of intensity, related to the color and texture, is not considered.
shows the grey intensity range for surfaces I–IV. Some of the intensity ranges are overlapping, for example, surfaces II and III, which preclude the use of a single threshold to segment the grey intensity.
For the three experiments conducted in this study, the surfaces are 25 m, 5 m, and 6 m away from the ILRIS3D (Figure 1
). All of the experiment results show similar patterns of concentric circles and distinct linear features at the boundaries of two surfaces (Figure 3
). It is expected that this pattern of the intensity data can be found in other dataset of surfaces collected by ILRIS3D at greater distance.