Professional Documents
Culture Documents
A time-of-flight camera suffers from significant range distortion due to the scattering artifact caused by secondary reflections occurring between the lens and the image plane. The reflected beam from the foreground objects undergoes multiple reflections within the camera device thereby introducing parasitic signals that bias the late-arrival, backscattered signals from the background targets. These additive signals cause degradation of the depth measurement for the farther objects thus limiting the use of range imaging camera for high precision close-range photogrammetric applications. So far the modelling of scattering distortion has been limited to base on the plausible linear system model using inverse filtering approach. The complicated nature of multiple internal reflections of the signals between the lens and the imager unfavourably hinders the use of linear system model for all scattering environment. A simple planar foreground scattering surface infront of another planar surface invalids the use of deconvolution compensation method for reversing the scattering artefact. In the absence of strong physical basis, the only alternative is to heuristically quantify the range distortions through exhaustive experimentations. A simple empirical modelling of range bias due to scattering distortion using analytical curve-fitting method is presented herein. Keywords: Range camera, Scattering, Empirical Modeling, Spline Interpolation, Curve-fitting, Surface-fitting
2 Introduction
Three-dimensional range imaging camera systems are a recent development for close-range terrestrial photogrammetric applications. They operate based on the phase-shift principle to determine the distance between the target and the camera. Each pixel in the sensor frame independently measures distance and amplitude information of the scene which is realized through CCD/CMOS lock-in pixel technology (Lange and Seitz, 2001). Unlike 3D laser scanners, a range camera does not need to sequentially scan its field of view to collect spatial and radiometric information. The range and the amplitude information are obtained simultaneously by sampling the returned modulated optical signal at every element location of the solid-sate sensor. Like other measuring devices, the range cameras also suffer from geometric and radiometric distortions which need to be accounted for through calibration process. The methodology of camera calibration is well established for all 2D imaging cameras, however, the 3D range imaging cameras are not able to be calibrated efficiently using traditional approach due to complicated systematic biases such as scattering effect on the range measurements. Research is
underway for developing a calibration procedure for the range cameras by incorporating range measurements in a self-calibration approach (Lichti, 2008) or by separately modelling the range distortions beside performing standard digital camera calibration (Jaakola et al, 2008; Kahlmann and Ingensand, 2007). On the contrary, only few attempts are being made to quantify the scattering effect caused by the multiple signal attenuation. Kavli et al (2009) and Dubois et al (2007) have published few results on compensation of the scattering bias using inverse filtering approach, where they basically use trial and error method of defining the inverse filter based on Gaussian or empirically defined point-spread function (PSF) approximation. Nonetheless, the linear system model presented by them is questionable due to complicated and uneven nature of the occurrence of multiple internal reflections inside the camera system. Therefore in the absence of strong physical basis, the only alternative is to empirically formulate the range distortions through exhaustive experimentations. (Kahlmann, 2006; Dubois et al., 2007; Kavli et al., 2009)
Section 1 presents a brief description of the nature of complications of multiple internal reflections that is occuring within the camera device. Section 2 shows the results of scattering distortions as a function of different parameters. Section 3 outlines empirical method of compensating scattering distortion using analytical curve-fitting method. Section 4 presents results from SR4000 range camera.
The angular phase-shift for the closer objects will be lower than the phase-shift for the longer object as shown in Figure 1, where A,B and C are the angular phase offsets, and the are the corresponding ranges of the point Q, Q and O respectively.
The point Q is the displaced point Q when a brighter closer object O is present in the imaging scene which causes scattering range effect on the background object point due to multiple internal reflection of the signals within the camera system as shown in Figure 2. The multiple internal reflections of the early arrived signals from the foregreound objects attenuates the weak signal from the background object thus lowering its angular phase offset causing shortened range measurement. The exact nature of multiple internal reflection of the signals is difficult to describe with a physical model because of the scattering phenomenon is highly imaging scene variant. Figure 2 hows a schematic representation of how multiple internal light reflections from the closer object could attenuate the actual signal from the far object. The difficulty of measuring the signal attenuation by unknown number of internal reflections poses limitation to any perceived physical model of the scattering artefact. For instance point P can be attenuated with multiple signals due to internal reflections.
The solution to this problem explicitly requires modelling of the scattering PSF ( of the camera so that a method of deconvolution can be employed to undo the effect of the scattering. They used a linear shift-invariant system model to quantify the scattering bias using a blinddeconvolution approach where the point-spread function of the camera including scattering bias is plausibly defined by a trial and error method using a Gaussian approximation. This method is limited due to the non-idealization of point source of light from the camera for measuring the point-spread function, which is fundamental in the linear-system model. Kavli et al. (2009) uses the same approach of linear system model to compensate for the scattering distortions in ToF cameras using generally shaped empirical models for the point spread function. Nonetheless, the linear system model presented by them is questionable due to complicated nature of the occurrence of multiple internal reflections inside the camera system. An edgespread function (ESF) experiment using two planar objects separated at certain distance invalids the use of linear system model for compensating scattering distortion. Figure 3 show the intensity
image of the scattering scene where a foreground object (a plane board) is placed one and half meters away from the background object (a wall).
The superimposed image point clouds of the scattering experiment scene of with and without the presence of the scattering object (a foreground planar board) are shown in Figure 4. The linear band of points at the upper portion in Figure 4 is the image point clouds of the object scene with the background wall. The stepped-like band of points is the point clouds of the object scene with a planar board placed in front of the wall, with about fifty percent of each surface appearing in the sensor plane.
It is clearly visible in Figure 4, the displacement of the background wall towards the camera when another object (a plane board) is introduced in the imaging scene. This is caused by the scattering phenomenon. Dubois et al. (2007) and Kavli et al. (2009) defined scattering compensation model with a physical model based on the linear shift-invariant system. The linear system model typically requires defining or measuring the point-spread function (PSF) accurately in order to successfully undo the filtering operation. However, direct measurement of the scattering PSF is impossible owing to the non-idealization of point source of light from the camera. Often line spread function (LSF) or edge spread function (ESF) is measured to deduce PSF indirectly. Figure 5 shows the superimposed theoretical and measured ESF obtained from the scattering experiment. The dotted line represents expected theoretical ESF, whereas the solid line represents the measured ESF obtained by fitting a curve on one row pixel from the step-liked band of points shown in Figure 4.
-300
-200
-100 X (mm)
100
200
Preliminary results from the ESF experiments indicated that the measured ESF is not conforming to the expected theoretical ESF of the scattering distortion profile. This ESF anomaly questions the relevance of the linear system model for the scattering phenomenon portraying in the 3D range camera. The scattering phenomenon is a complicated event due to multiple reflections occurring inside the camera which may not be correctly described by the linear system model. In the absence of concrete physical scattering model, it is imperative to explore the empirical methods of modelling scattering distortion through exhaustive experimentations. This paper presents an empirical modelling of range bias due to scattering distortion using analytical curvefitting method. (Note: Figure 4 and 5 can be placed together in Matlab).
because of the unavailability of a large foreground planar object which is required to cover whole sensor frame by the foreground object. At each location, multiple images were taken to reduce the random noise of the system. Images were captured at four different integration times at 10 ms, 24 ms, 36 ms and 51.2 ms for all scenes. Figure 6 shows the experimental set up for the scattering experiment. A white projector screen of size 96 inch square was placed in front of the background wall surface, which acts as a scattering object.
The scattering artefact in SR3100 range camera is portrayed in the form of range and amplitude bias on the background objects. The measured range and amplitude for the background objects in presence of scattering foreground object is lower than the actual measurements without the scattering object. Following figures clearly shows the effect of scattering artefact of the range camera. Figure 7 shows the range and amplitude bias of one row of pixels as a function of the surface area of the scattering object (camera distance need to be noted). The camera and the background object are fixed at 380 cm. The scattering object is laterally moved at ten percent increment located at a distance of 140 cm from the camera. Scattering edge is defined as the location of the edge of the foreground object where it overlaps with the background object. The
trend is clearly visible where the range and amplitude bias are increasing with the increase in the surface area of the scattering object. Additionally, the range and amplitude bias have a clear linear relationship which suggests proportional change in the phase shift of the attenuated signal as a result of the amplitude attenuation caused by the scattering artefact.
600 500 400 90 % 80 % 70 % 60 % 50 % 40 % 30 % 20 % 10 % 300 250 200 90 % 80 % 70 % 60 % 50 % 40 % 30 % 20 % 10 %
Amplitude bias
8
4 X (mm)
4 X (mm)
Figure 7: Range and amplitude bias as a function of the surface area of the scattering object
600
Amplitude bias
2 X (mm)
2 X (mm)
Figure 8: Scattering induced bias when camera and scattering object is at 300 and 200 cm from the background object
Amplitude bias
600
600 400
400
2 X (mm)
2 X (mm)
Figure 9: Scattering induced bias when camera and scattering object is at 380 and 200 cm from the background object
Figures 8 and 9 show the range and amplitude bias as a function of the integration time used for the data capture. All four different integration time for different scattering scene environment produces almost same random range bias. This suggests that scattering effect on range measurements is invariant of the integration time. However the apmlitude bias due to scattering is dependent on the integration time. The monotonicity relationship between the integration time used and the scattering induced amplitude bias is clearly observed which is because of the more number of photons impinging on the sensor receptor at higher integration time thus recording high amplitude value.
300 250 200 300 cm 260 cm 220 cm 180 cm 140 cm 500 Scattering Edge 300 cm 260 cm 220 cm 180 cm 140 cm Scattering Edge
400
Amplitude bias
0 1 2 X (mm) 3 4 5
200
100
-100
2 X (mm)
Figure 10: Range and amplitude bias as a function of the distance between the camera and the scattering object
Figure 10 portrays range and amplitude bias of one central row of pixels as a function of the distance of the scattering object from the camera. The camera and the background object are fixed at 380 cm. The scattering object is moved at 40 centimeters increment from 140 to 300 centimenetrs from the camera. It is observed that the range and amplitude bias increases with the increasing distance between the background and the scattering objects. This is expected because
the power of the signals decays as an inverse square of the distance. When scattering object is closer to the camera relative to the background object, the reflected light from the scattering object has more power than the reflected light from the farther object causing greater signal attenuation resulting into proportional scattering bias.
Figure 11: Range bias when camera at 140cm and 380 cm from the scattering and background object respectively
Figure 12: Range bias when camera at 240cm and 380 cm from the scattering and background object respectively
Range bias due to scattering is the subject of interest in this study. Dubois et al. (2007) and kavli et al. (2009) reported a maximum of 40 centimeters of range bias due to scattering based on their experiment. On the contrary, this study has shown that the scattering induced range bias could reach up to 250 centimeters in presence of a highly reflective large surface area scattering object when the scattering and the background objects are separated at appreciable distance as shown in Figure 11, left image. Figures 11 and 12 show the variation in range bias at different scattering scene environment. In both the figures, it is clearly visible that the scattering induced range bias
is more in the periphery than the inside portion of the image plane. This is due to the greater power loss of the reflected signal at the periphery than the middle portion of the target scene. It has been reported by Jaakola et al. (2008) on the additional power loss of the SR3100 range camera at the periphery besides the cosine-fourth power loss observed in the standard optical system.
In Figures 7 to 12, it is observed that the scattering induced bias is greater than the inherent system noise as depicted in Figure 13, where twenty consecutive images was used to evaluate the range and amplitude biases of the fixed target scene. The range noise is within 400 mm and amplitude bias is within 300 16-bit quantized values.
4.2.2
Chiabrando et al. (2009) reported absence of scattering distortions for SR4000 range camera which is a fourth generation range camera after SR3100. Similarly, this study also did not observe any significant range or amplitude bias unlike the scattering induced distortions observed in the previous generation range camera. Figures 14 and 15 shows the scattering induced range and amplirtude bias for SR4000 range camera which proves that the scattering artefact is greatly reduced or eliminated as they are within the range of the noise of the consecutive images of the range camera shown in Figure 15. Tewnty consecutive images were used to evaluate the differences of range and amplitude measurements of the same target scene for the two images taken consecutively. The range noise is within 100 mm and amplitude noise is within 1000 16-bit quantized values. It is however unknown how the scattering artefact in SR4000 is rectified or atleast minimised within the noise of the system by the manufacturer, either through software implementation or hardware consolidation.
Amplitude bias
50 % 40 % 30 % 20 % 10 %
4 X (mm)
-400
4 X (mm)
Figure 14: Range and amplitude bias as a function of the surface area of the scattering object for SR4000
30 300 cm 210 cm 160 cm 500 400 300 200 100 0 -100 300 cm 210 cm 160 cm
20
Scattering Edge
10
-10
-20
2 X (mm)
Amplitude bias
Scattering Edge
2 X (mm)
Figure 15: Range and amplitude bias as a function of the distances of the scattering object from the camera for SR4000
5 Methodology
5.1 Smoothing: 3D Surface Fitting
A smoothing ridge estimator is used to fit the 3D surface on a 2D grided data points. The elevation at the unknown data points are estimated using a linear interpolation on a triangular mesh. Equation 2 gives the linear interpolation method for a traingulated surface depicted in Figure 17.
The smoothing is achieved by a spring-beam approximation where the data points are connected to the flexible beam by springs. A flexible beam is gernerally modelled with a cubic spline where the minimization function is the potential energy stored in the extended springs which is due to both extension of the springs and the bending of the beam. The potential energy in the spring is equal to the square of the length of the extended spring which is equivalent to the L2 norm. But the L2 norm is variable in accordance to the smoothing constraint. Thus the ridge estimator used for fitting a homeomorphic triangular mesh is biased towards smoothing which is achieved by weighting the L2 norm. Figure 18 shows the schematic description of the spring-beam approximation on a scattered discrete data points.
An approximation of the 3D surface greatly reduces the noise of the system as shown in Figure 19. The proposed scattering compensation model is based on the smoothed values of the scattering induced range bias.
1200 1000
1000 800
Constraining Equation 3 with the following four conditions gives a unique solution. the spline pass through all data points first derivative is continuous at all interior points second derivative is continuous at all interior points boundary conditions at end points are known When the boundary conditions at the end points are defined as in Equation 4, the spline is called Not-a-Knot Spline.
When the first derivatives at the end points are fixed to a scalar values then it is called clamped spline. Not-a-Knot spline is chosen over the clamped spline because it fitted the data points better than the clamped spline as shown in Figure 20.
Spline Fit 700 600 Data Points Not-A-Knot Spline Clamped or Complete Spline
20
70
Figure 21: Flow diagram of the two different scattering compensation models
2500
Range bias(mm)
150
2000
Count
4 2 Y(mm) 0 0 2 X(mm) 4
100 50
1500
1000 0 6 500
0 -40
Figure 22: Camera and scattering object at 380 and 160 cm respectively from the background object
Superimposed of actual and corrected surface 4000 % Top: Actual surface 600 % Bottom: Interpolated surface 3500 3000 2500
Success rate of scattering compensation % Over Compensated < 0 % Under Compensated > 0
Range bias(mm)
500 400
Count
4 2 Y(mm) 0 0 2 X(mm) 4
-10
-5 0 5 Percent Compensated
10
15
Figure 23: Camera and scattering object at 380 and 200 cm respectively from the background object
Superimposed of real and interpolated range bias % Top: Interpolated surface % Bottom: Actual surface 250
Success rate of scattering compensation 3500 3000 2500 2000 1500 1000 4 500 0 -100 % Over Compensated < 0 % Under Compensated > 0
Range bias(mm)
Count
50
Figure 24: Scattering compensation approximated for scattering obeject at 260 cm from the camera using all the dataset