WO2016156308A1

WO2016156308A1 - Method and system for real-time motion artifact handling and noise removal for tof sensor images

Info

Publication number: WO2016156308A1
Application number: PCT/EP2016/056788
Authority: WO
Inventors: Cedric SCHOCKAERT; Frederic Garcia Becerro; Bruno Mirbach
Original assignee: IEE International Electronics and Engineering SA
Current assignee: IEE International Electronics and Engineering SA
Priority date: 2015-04-01
Filing date: 2016-03-29
Publication date: 2016-10-06
Anticipated expiration: 2017-10-01
Also published as: CN107743638B; US11215700B2; US20180067197A1; CN107743638A; DE112016001560T5; LU92688B1

Abstract

A method and system for real-time motion artifact handling and noise removal for time-of-flight (ToF) sensor images, the ToF sensor images corresponding to received signals (r(t)) received by a ToF sensor following sending of modulated sent signals (s(t)). The method comprises: calculating values of a cross correlation function C(T) at a plurality of temporally spaced positions or phases from the sent (s(t)) and received (r(t)) signals, thereby deriving a plurality of respective cross correlation values [C(T₀), c(T₁), C(T₂), C(T₃)]; deriving, from the plurality of cross correlation values [C(T₀), c(T₁), C(T₂), C(T₃)], a depth map D, the depth map D having values representing, for each pixel, distance to a portion of an object upon which the sent signals (s(t)) are incident; deriving, from the plurality of cross correlation values [c(T0), c(Ti), c(T2), c(T-i)], a guidance image (/; /'), the guidance image (/; /') being an image having well defined edges; and generating an output image D' based on the depth map D and the guidance image (/; /'), the output image D' comprising an edge-preserving and smoothed version of depth map D, the edgepreserving being from guidance image (/; /'). The method may further comprise: filtering the guidance image / to generate a de-noised guidance image t, and generating an output image D' based on the depth map D and the de-noised guidance image f. The method may further comprise generating a plausibility map P based on the plurality of cross correlation values [C(T₀), C(T₁), C(T₂), C(T₃)], the plausibility map P comprising, for each pixel of the depth map D, a value indicative of whether the pixel is corrupted by motion or saturation; and generating the output image D' based on the depth map D and the plausibility map P, and on either the guidance image / or the de-noised guidance image f.

Description

Method and System for Real-Time Motion Artifact Handling and Noise Removal for TOF Sensor Images

Technical field

[0001 ] The present invention relates to image processing, and more particularly to a method and system for real-time motion artifact handling and noise removal for Time-of-Flight (ToF) sensor images.

Background Art

[0002] There are a range fields and applications in which accurate distance calculation by means of a ToF sensor (e.g. camera) is required. Those scenes captured by the camera may be either static or dynamic, and in the latter case the images may be corrupted by motion artifacts occurring at moving object boundaries.

[0003] Typical applications involving accurate depth measurements are human- computer interaction, in which fast hand segmentation for gesture recognition is sought, and automotive related applications, where occupancy classification, head pose estimation and/or environment perception is to be performed. Other applications will be apparent to the skilled person.

[0004] ToF cameras based on demodulation lock-in pixels estimate depth from the phase-shift between emitted and received modulated near-infrared (NIR) signals, in which four sequential phase-shifted images are required, i.e., the "four- taps" technique.

[0005] The conventional ToF working principle assumes the scene to be motionless during this time interval. However and in practise, unreliable depth measurements arise along object boundaries in dynamic scenes, especially when fast movements are involved.

[0006] Several approaches have been already proposed to deal with the motion artefact downside of TOF cameras. They are mainly classified in two main categories: (1 ) pixel-based and (2) flow-calculation-based correction. [0007] In Hussmann, S., Hermanski, A., & Edeler, T. (2010), "Real-time motion suppression in TOF range images", in IEEE Instrumentation & Measurement Technology Conference Proceedings, pp. 697-701 , 2010, the authors introduced a method to compensate the linear motion of objects on a conveyor belt. The linear shift is measured and compensated on the four phase-shifted images before depth calculation. However, their approach is restricted to linear motion within the range of the camera field-of-view.

[0008] Schmidt, M. "Analysis Modeling and Dynamic Optimization of 3D Time-of- Flight Imaging Systems", PhD Thesis, Natural Sciences and for Mathematics of the Ruperto-Carola University of Heidelberg, Germany, 201 1 , proposed a method to detect and correct the motion per pixel by analysing the temporal relationship along a set of N times four phase-shifted images acquired in burst mode; N being the size of the set. The method is constrained by the ToF camera hardware specifications, which must provide the N times the four phase-shifted images set.

[0009] More recently, Lee, S. "Time-of-Flight Depth Camera Motion Blur Detection and Deblurring", Signal Processing Letters, IEEE, 21 (6), 663-666, 2014, has proposed a new pixel-based correction. The author developed several metrics that permit correction of pairs of corrupted phase-shifted images. However, this method is only applicable to those ToF cameras for which the integration time is divided in two stages, i.e., phase-shifted image pairs are recorded simultaneously.

[0010] In Hoegg, T., Lefloch, D., & Kolb, "A. Real-Time Motion Artifact Compensation for PMD-ToF Images", Lecture Notes in Computer Science, pp. 273-288, 2013, the authors compensate the linear motion per pixel using an estimated flow field generated from the four phase-shifted images. Similarly, in Lindner, M., & Kolb, A. "Compensation of Motion Artifacts for Time-of-Flight Cameras", in Lecture Notes in Computer Science, Vol. 5742, pp. 16-27, 2009, optical flow is also computed between consecutive phase-shifted images in order to compensate the motion before depth calculation. Consequently, the method requires three optical flow calculations which makes the method non practical for real-time application.

[001 1 ] In Lefloch, D., Hoegg, T., Kolb, A.: "Real-time motion artifacts compensation of ToF sensors data on GPU", in Proceedings of SPIE Vol. 8738, 2013, the authors reduced the complexity of the previous method by replacing the third optical flow by a polynomial approximation. However, real-time performance remains an issue.

Technical problem

[0012] A problem with known methods is how to address the motion artefact due to moving objects in the scene, in particular how achieve robust detection of corrupted depth pixels due to motion artefact or saturation.

[0013] A further problem is how to achieve real-time performance, enabling for further processing, e.g., gesture recognition or head pose estimation.

[0014] A further problem with known methods is how to achieve noise removal whilst preserving edges.

[0015] A further problem is how to select a guidance image to reconstruct the boundaries of moving objects.

[0016] A further problem with known methods is how to achieve noise and texture removal from the guidance image.

[0017] Therefore, an effective real-time approach to address the motion artefact related to the ToF camera's working principle is needed.

[0018] It is an object of the present invention to provide a system and method that address at least some of the foregoing problems and provide improved techniques for real-time motion artifact handling and noise removal for ToF sensor images.

Summary of the Invention

[0019] According to one aspect of the invention there is provided a method for real-time motion artifact handling and noise removal for time-of-flight (ToF) sensor images, the ToF sensor images corresponding to received signals (r(t)) received by a ToF sensor following sending of modulated sent signals (s(t)), the method comprising: calculating values of a cross correlation function C(T) at a plurality of temporally spaced positions or phases from the sent (s t) ) and received (r(t)) signals, thereby deriving a plurality of respective cross correlation values [C(T₀), C(T₁), C(T₂), C(T₃)] ; deriving, from the plurality of cross correlation values [ a depth map D, the depth map D having values representing, for each pixel, distance to a portion of an object upon which the sent signals (s(t)) are incident; deriving, from the plurality of cross correlation values

), ₃)], a guidance image (I; I'); and generating an output image D' based on the depth map D and the guidance image (I; I'). Preferably, the guidance image (/; /') is an image having well defined edges. Preferably, the output image D' comprises an edge-preserving and smoothed version of depth map D, the edge-preserving being from guidance image (/; /')■

[0020] The method may comprise acquiring a plurality of phase shifted images in succession, each phase shifted image corresponding to a respective temporally spaced position or phase. The phase-shifted images may comprise amplitude images.

[0021 ] Each cross correlation value [c(T₀), c(T_i), c(T₂), c(T-_i)] may correspond to one of a plurality of phase-shifted image, and the guidance image (/; /') is generated from the phase-shifted images.

[0022] The step of deriving the guidance image comprises selecting as the guidance image a phase-shifted image from a plurality of previously-acquired phase-shifted images, based on one or more predetermined criteria. The predetermined criteria may comprise that the phase-shifted image (i) with maximum amplitude of the object degraded by motion artefact, (ii) with maximum object edge sharpness value, (iii) with the best edge contrast and/or image SNR, (iv) with the maximum average spatial amplitude or (v) that is non-corrupted, is selected as the guidance image.

[0023] The method may include using a guided filter (GF) to apply valid depth measurements to previously identified corrupted depth pixels due to motion artifacts.

[0024] The step of generating an output image D' may comprise determining the output image D' as:

[0025] In one embodiment, generating an output image D' comprises: filtering the guidance image / to generate a de-noised guidance image f , and generating an output image D' based on the depth map D and the de-noised guidance image t. The step of filtering the guidance image / to generate a de-noised guidance image t may comprise using a guided filter to perform said filtering.

[0026] The step of generating an output image D' may further comprise: generating a plausibility map P based on the plurality of cross correlation values the plausibility map P comprising, for each pixel of the depth map D, a value indicative of whether the pixel is corrupted by motion or saturation; and generating the output image D' based on the depth map D and the plausibility map P, and on either the guidance image / or the de-noised guidance image /. The step of generating the plausibility map P may comprise determining, for each pixel, a metric p_t representing the deviation of the phase-shifted amplitudes from a sinusoidal model of the cross-correlation function. The metric p_t may comprise:

where a is a regularization parameter preventing high value of p_t when the amplitude a is low. The plausibility map P may have values P_t, for each pixel /^' such that:

if metric Pi > δ

otherwise

where δ is a motion threshold value, above which motion is detected at pixel /^'. The motion threshold value δ may be determined by capturing by the ToF sensor of an empty or motionless scene.

[0027] In one embodiment, filtering the guidance image / to derive the de-noised guidance image / comprises: applying an edge preserving de-noising filter to guidance image /, whereby edge information and noise reduction from the guidance image / is transferred to the output image D'.

[0028] In one embodiment, filtering the guidance image / comprises deriving de- noised guidance image / using:

where «i = ]^|∑fcew_t ctfc. ^{witn a}k = -

*Ί = A∑fcew_t h, with b_k = (1 -a_k).F_k, σ/(² is the variance of / in w_k,

\w\ is the number of pixels in w_k,

ε is a regularization parameter penalizing large a_/r, and h = ^∑i_€w_k is the mean of I in w_k.

[0029] In one embodiment, generating an output image D' comprises generating an output image D' according to:

Ό = a l + b_v

where

D_k = -— :∑_iewt DiPi is the mean of D in w_k weighted by the map P,

|w| is the constant number of pixels in the window w_t centered at pixel i,

\^w _k\ = ∑ie _wfc Pi is the number of pixels in w_k where the plausibility is 1 , and

°l= ^∑_iewfc /;²P;-7^where _fc= ^∑_iewfc

[0030] The output image D' may comprise a version of depth map D alternatively or additionally from which motion artifacts and/or noise have been suppressed or removed.

[0031] The cross correlation function C(T) may be calculated as:

where the s(t) is the sent signal given by s(t) = 1 + cos(wt),

r(t) is the received signal given by r(t) = h + a. cos(wt - ψ) ω = 2nf_m is the angular modulation frequency,

f_m is the modulation frequency, and

h is the background light plus the non-modulated part of the sent signal.

[0032] The cross correlation values [c(T₀), c(T_i), c(T₂), c(T-_i)] may be determined from C(T) as four samples (r_k), k = 0, . . ,3 , taken at four subsequent time intervals T_K = k. - = k/4f_m within a modulated period T. In one embodiment, distance measurements d for each pixel of the depth map D are obtained from

c = 3. 10⁸ m/s, the speed of light,

L is the working range or non-ambiguity distance range of the ToF sensor, and L =—.

2/m

[0033] The method may further comprise deriving an amplitude image A defined as A = Γα,-,-1 , where the , , are determined from: a = 1/2 (C(T₃) - c(_Tl))² + (C(T₀) - C(T₂))² ,

where C(T₀), c^r^, C T₂), C(T₃) are the cross correlation values.

[0034] In one embodiment, four cross correlation values [C(T₀), C(T₁), C(T₂), C(T₃)] are calculated from the cross correlation function C(T) .

[0035] According to another aspect of the invention there is provided a programmable image processing system when suitably programmed for carrying out the method of any of the preceding claims, the system comprising circuitry for receiving or storing the received signals (r(t)) and sent signals (s(t)), and processing circuitry for performing the methods.

[0036] According to another aspect of the invention there is provided a system for real-time motion artifact handling and noise removal for time-of-flight (ToF) sensor images, the ToF sensor images corresponding to received signals (r(t)) received by a ToF sensor following sending of modulated sent signals (s(t)), the system comprising: circuitry for receiving or storing the received signals (r(t)) and sent signals (s(t)); processing circuitry, coupled to the circuitry for receiving or storing the received signals (r(t)) and sent signals (s(t)), the processing circuitry being operable for calculating values of a cross correlation function at a plurality of temporally spaced positions or phases from the sent (s t)) and received (r(t)) signals, thereby deriving a plurality of respective cross correlation values [ deriving, from the plurality of cross correlation values [ a depth map D, the depth map D having values representing, for each pixel, distance to a portion of an object upon which the sent signals (s(t)) are incident; deriving, from the plurality of cross correlation values

), ₃)], a guidance imagef/; I'), the guidance image (I; I') being an image having well defined edges; and generating an output image D' based on the depth map D and the guidance image(/; /'), the output image D' comprising an edge-preserving and smoothed version of depth map D, the edge-preserving being from guidance image (/; /')■

[0037] According to another aspect of the invention there is provided a recordable, rewritable or storable medium having recorded or stored thereon data defining or transformable into instructions for execution by processing circuitry and corresponding to at least the steps of any of claims 1 to 24 of the appended claims.

[0038] According to another aspect of the invention there is provided a server computer incorporating a communications device and a memory device and being adapted for transmission on demand or otherwise of data defining or transformable into instructions for execution by processing circuitry and corresponding to at least the steps of any of claims 1 to 24 of the appended claims.

[0039] In each of the above aspects and embodiments, the emitted/sent and received signals may comprise modulated near-infrared (NIR) signals.

[0040] In certain embodiments, a robust method to identify those pixels in the resulting depth map that are prominent such as to be unreliable is provided. Then, their values are replaced by the closest reliable ones using the guided filter (GF) and an accurate guidance image generated from the previously acquired sequential phase-shifted images.

[0041 ] As used in embodiments, the GF presents a better behaviour near edges than alternative edge preserving filters such as the widely used bilateral filter (BF), with a major advantage of being a fast and non-approximate linear time algorithm.

[0042] The present invention satisfactorily addresses the motion artefact, even in extreme conditions.

[0043] In contrast to the abovementioned prior art motion artefact compensation methods, the techniques disclosed herein only require a non-corrupted phase- shifted image (considered to be the guidance image), regardless of the ToF camera architecture and the way the distance is calculated. In addition, a real-time performance is achieved without parallelized computation on a GPU.

[0044] An advantage of the invention, at least in embodiments, is in addressing the motion artifact from depth map in real time. The object shape is extracted from one of the 4 phase-shifted amplitudes selected according to a criteria taking into account, e.g., the image SNR and edges contrast. This phase-shifted amplitude is used as guidance image for an edge preserving de-noising filter in order to reconstruct the depth of the objects degraded by motion artifact. In the prior art, motion correction is either

a. performed pixel per pixel, without taking into account spatial correlation of the depth map, and usually creating other artifact like noise increase when the depth is corrected, or

b. corrected by spatial analysis using pixel correspondence method between the four phase-shifted amplitudes, but not compatible with real-time application (dense optical flow based methods).

[0045] A further advantage, at least in embodiments, is in edge preserving spatial de-noising of depth map. The guidance image is filtered by means of a dedicated edge preserving de-noising filter. The noise reduction of the guidance image is transferred to the depth map in the motion artifact suppression stage. In the prior art, spatial de-noising is applied to the depth map directly. A spatial de-noising of the depth map requires ideally the depth map to be cleaned from artifacts (motion artifact, flying pixel artifact,...) before being de-noised by an edge preserving filter to avoid the risk of spatial artifact propagation. In the proposed method, the depth map is de-noised by transferring the reduction of noise from the selected processed phase-shifted amplitude image. Phase-shifted amplitudes do not suffer from artifact like the depth map does, and as a consequence, object edges are usually better defined (no flying pixels).

[0046] A further advantage, at least in embodiments, is in the selection of a guidance image to reconstruct the border of moving objects in the depth map. A guidance image is defined as one of the 4 phase-shifted amplitudes. From the 4 phase-shifted amplitude images, the best guidance image is selected using a quality criterion, or a combination of criteria, taking into account, at minimum, the edge contrast and image SNR.

[0047] A further advantage, at least in embodiments, is in the process of the guidance image to reduce noise and remove texture. This image requires a dedicated processing

a. to remove unwanted artifact that could be transferred to the corrected depth map, and

b. to reduce the noise while object boundaries are preserved.

[0048] A further advantage, at least in embodiments, is in the detection of pixels corrupted by motion or saturation. A metric is used to measure the plausibility of each pixel of the depth map to be corrupted by motion artifact or saturation.

[0049] A further advantage is that, at least in embodiments, the invention enables an increase the integration time, i.e., the time period to acquire the 4 phase-shifted images in order to increase depth accuracy. This in turn might induce to motion artefact that might be handled by the method.

[0050] A further advantage is that, at least in embodiments, the invention also applies to depth accuracy increasing methods based on time-filtering, i.e., methods that combine several depth maps in order to increase depth accuracy. Non-matched depth pixels between the combined depth maps (due to moving objects) result in depth motion artefact that can be also addressed by the use of embodiments of the invention. Brief Description of the Drawings

[0051 ] Preferred embodiments of the invention will now be described, by way of example, with reference to the accompanying drawings, in which:

[0052] Figure 1 (PRIOR ART) shows (a) a schematic diagram of object illumination and image capture, (b) plots of modulated signal strength for sent (s(t)) and received (r(t)) signals, and (c) sampling at four instants of the cross correlation c(-r) function calculated from the sent (s(t)) and received (r(t)) signals, according to the known working principle of ToF cameras;

[0053] Figure 2 schematically illustrates the architecture for processing signals, obtained as illustrated in Fig. , in accordance with an embodiment of the present invention;

[0054] Figure 3 shows (a) the raw guidance image /, and (b) the processed (de- noised) guidance image /', derived using the system of Fig. 2; and

[0055] Figure 4 shows (a) the input depth map D, and (b) the processed (resulting) depth map D', derived using the system of Fig. 2.

Description of Preferred Embodiments

[0056] As used herein, the "images" or "image signals" may be analog or digital, and may be subject to conventional analog or digital filtering.

[0057] Where references are made herein to steps, operations or manipulations involving "images", etc., these are implemented, where appropriate, by means of operations upon electronically processable representations (e.g. captured stills of video frame signals, bitstream video data, MPEG files or video streams, PC-video, or any other capturable or viewable image data format) of such "images".

[0058] Where references are made herein to steps, operations or manipulations involving "images", "image signals" or "image data", these may be implemented, where appropriate, by means of software controlled processor operations, hardware circuitry or any suitable combination of these.

[0059] While the present invention is suitably embodied in a computer system, it may be incorporated in an adaptor, an image processor, or any other equipment located between or incorporating an image source or image capture device and a display device (e.g. LCD, Plasma, projector, etc.), or in the display device itself. The computer system suitably comprises a processor coupled (where appropriate via DACs and ADCs, or other interfaces) to RAM, ROM, storage devices, image capture and/or image storage devices, display driver and display devices, data communication and other peripherals, as is well known to persons skilled in the art; therefore, these will not be illustrated or discussed further.

[0060] In the following, the ToF working principle is briefly discussed, to facilitate the understanding of the present invention.

Time of Flight principle

[0061 ] Figure 1 (a) (PRIOR ART) shows a schematic diagram of object illumination and image capture, according to the known working principle of ToF cameras.

[0062] A ToF camera 102 includes a modulation element 104 generating a transmitted or sent signal s(t) that is emitted by optical emitter 106 as a modulated NIR illumination signal 108. The NIR illumination signal 108 is incident upon object 1 10 within a scene being sensed, with the reflected optical signal, reflected by object 1 10 being received at sensor (e.g.2D CCD array) 1 12 as received signal r(t).

[0063] Also within ToF camera 102, a phase delay element 1 14 receives the sent signal s(t) and applies a phase delay to it, thus outputting a phased delayed signal s(t+ r), where τ is a phase delay. Processing circuitry (not shown) within or coupled to sensor 1 12 then calculates, based on phased delayed signal s(t+ r) and received signal r(t), cross correlation function C(T) , as discussed in further detail below.

[0064] As illustrated in Fig. 1 (a), ToF cameras 102based on demodulation lock-in pixels provide distance measurements from the difference in phase between sent (s(t)) and received (r(t)) modulated near-infrared (NIR) signals. The amplitude and phase of the incoming modulated received (r(t)) is retrieved by synchronously demodulating the received signal (r(t)) within the sensor/detector. [0065] Figure 1 (b) (PRIOR ART) shows plots of modulated signal strength for sent (s(t)) and received (r(t)) signals, according to the known working principle of ToF cameras. Cross correlation between the received (modulated) signal r(t) of amplitude a and phase φ, and the emitted modulated signal s(t), is performed. The phase φ of the received signal (r(t)) can be determined by taking the measurement of the cross correlation function C(T) at selectively chosen temporal positions or phases. The signals s(t) and r(t) are assumed to be sinusoidal, and may be expressed

s(t) = 1 + cos(a>t)

r(t) = h + a. cos(a>t— φ)

with ω = 2nf_m the angular modulation frequency, with f_m the modulation frequency, and with h the background light plus the non-modulated part of the incident signal; the waveforms and their relationships are illustrated in Figure 1 (b).

[0066] The cross correlation function C(T) is calculated as follows:

C(T) = i J^__Q r(t). s(t + r)dt = h + ^ COS(O>T + φ).

[0067] Three or more samples of C(T) per modulated period T are usually needed in order to unambiguously determine the phase φ and the amplitude a of the incident signal, as well as its offset h. In embodiments disclosed herein, the so- called four-taps technique is used, in which four samples (τ_¾), k = 0, . . ,3, are taken at intervals r_k = k. - = k/4f_m within a modulated period T. Figure 1 (c)

(PRIOR ART) shows sampling at four instants of the cross correlation C(T) function calculated from the sent (s(t)) and received (r(t)) signals. As a result,

a = 1/2 V (C(T₃) - c(_Tl))² + (C(T₀) - C(T₂))²

C(T₀) + cOi) + C(T₂) + C(T₃)

= 4 ^

[0068] In embodiments disclosed herein, four samples instead of three are chosen, to (i) improve robustness against noise, (ii) enable a highly symmetric design of the sensor, (iii) ensure that the phase is insensitive to quadratic non- linearities in detection, and (iv) simplify the formulae for the phase φ, the amplitude a, and the offset h.

[0069] The working principle of ToF cameras 102 based on modulated NIR light resolves distance calculation from four phase-shifted images. Ideally, the four phase-shifted images would be simultaneously acquired, but in practice the acquisition is done sequentially. This in turn can cause corrupted distance calculations in those regions of non-matching raw phase values due to motion, that is, along object boundaries and within inhomogeneous reflection surfaces, which are more prominent the faster the object moves, the closer the object is to the ToF camera 102, and the higher the scene is exposed (higher integration time). Therefore, a larger integration time may be set for static scenes or scenes with slow moving objects, which would increase the depth accuracy, whereas and despite the increase of noise, short integration times may be set for high dynamic scenes with fast moving objects in order to avoid motion artefacts.

[0070] The distance measurements d to the object 1 10 in Fig. 1 (a) are obtained from

L

^{D =} Ϊ Ζπ - Φ with c = 3. 10⁸ m/s the speed of light and L the working range or non-ambiguity distance range of the ToF camera 102:

[0071 ] The factor 1/2 is due to the fact that light travels twice the distance between the camera 102 and the sensed object 1 10.

[0072] The ToF camera 102 incorporates, as will be appreciated by persons skilled in the art, an image sensor 1 12 whose size corresponds to the camera resolution (m x n). Hence, each single pixel constituting the image sensor 1 12 is identified by the pixel position (i, j), where /^' indicates the row and j indicates the column. Each pixel measures a distance d_i; to the object 1 10 (or a respective discrete portion thereof). As a result, the ToF camera 102 provides a distance image or depth map D defined as D = [d-ij] , the matrix of all the elements dij .

[0073] In the same way, an amplitude image A is defined as A = [¾^■] ■

[0074] Figure 2 schematically illustrates the architecture or system 200 for processing signals, obtained as illustrated in Fig. , in accordance with an embodiment of the present invention. The processing may involve any of the signal generation, mathematical processing and signal and parameter representation as describes above in relation to Fig. 1 , except where described otherwise in the following.

[0075] Briefly stated, from values [C(T₀), C(T₁), C(T₂), C(T₃)] of the correlation function the depth map D is derived by depth map module 204, the depth map D comprising values representing, for each pixel thereof, a distance to an object upon which the sent signals are incident. Also based on the values of the correlation function c(-r) , a guidance image / is generated by guidance image module 206, and, in a preferred embodiment, a de- noised guidance image /' is generated from guidance image / at guidance image de-noising module 208. (I in a preferred embodiment, however, the guidance image / may be used.) Finally, an output image (processed depth map) D' is generated and output by motion artifact handling module 21 0, based on depth map D and guidance image / or more preferably de-noised guidance image /'. In a further preferred embodiment, plausibility map generation module 212 generates a plausibility map P; and the output image (processed depth map) D' is generated and output by motion artifact handling module 21 0 based on (i) depth map D, (ii) plausibility map P and (iii) guidance image / or more preferably de-noised guidance image /'.

[0076] The processing by the various modules in Fig. 2 will be discussed in further detail below.

Guided filter

[0077] In this section, the guided filter (GF), employed embodiments of the invention, is briefly discussed: this is used to (1 ) de-noise the guidance image in de-noising module 208, and (2) set valid depth measurements to the previously identified corrupted depth pixels due to motion artifact.

[0078] The GF, in a preferred embodiment, is a new edge-preserving smoothing filter that, compared to the widely used bilateral filter, presents a better behavior near edges with a major advantage of being a fast and non-approximate linear time algorithm (0(/V) time), regardless of the kernel size and the intensity range.

[0079] Given a depth map D and a guidance image /, the resulting edge- preserving from / and smoothed version of D, i.e. D' is expressed as:

where

b_t = ^∑k _Wi b_k, with b_k = D_k-a_kl_k, are linear coefficients assumed to be constant in w_k. l_k and a_k are respectively the mean and the variance of / in w_k, \w\ is the number of pixels in w_k and ε is a regularization parameter penalizing large a_k. D_k = τ^-∑ί _{e Wt} is the mean of D in

I I ^

Wk -

[0080] The selection of the window size w_k may be done according to application: it may be small for image detail enhancement in order to enhance thin details, and larger for structure transferring filtering. The smoothing level is given by the ε parameter.

Plausibility map

[0081 ] Each pixel of the four phase-shifted images acquired for distance calculation are samples [C(T₀), C(T₁), C(T₂), C(T₃)] of the cross correlation function C(T) between the emitted s(t) and received r(t) sinusoidally modulated signal, as illustrated in Fig. 1 . Hence, if motion occurs during that acquisition time period, the temporal dependency of the cross correlation function C(T) will invalidate its assumed sinusoidal nature. Indeed, each sample will be located on a different sinusoidal curve, cancelling the correlation between samples. [0082] According to an embodiment, a pixel /^' affected by motion is identified by the following metric

Pi = kOi) - C(T₀) - C(T₂) + c(T₃) | /(a + a) where a is a regularization parameter preventing high value of p_t when the amplitude a is low.

[0083] In this embodiment, a motion is detected at pixel /^' if its plausibility is larger than a threshold δ:

otherwise

with δ a motion threshold value. The value of motion threshold value δ may be easily derived or adjusted by recording an empty or motionless scene by the ToF camera 102.

Guidance image selection and processing

[0084] A guidance image / with well-defined and sharp edges is needed to adjust the object boundaries in D affected by the motion artefact. Selection of the guidance image is performed in guidance image module 206 in Fig. 2.

[0085] If it is assumed herein that the motion during each phase-shifted image c(Ti) acquisition is negligible, any of the four phase-shifted images could be considered as a guidance image. However, as each phase-shifted image corresponds to a sampling of the cross correlation function C(T) between the received (r(t)) and emitted (s(t)) modulated signals, the phase-shifted image having the maximum intensity will have the best SNR and thus, the best contrast at object boundaries. Therefore, in an embodiment, the phase-shifted image having the maximum average spatial amplitude is then selected as a guidance image /. A further step is preferably done in order to avoid transferring the noise from / to the filtered D. That is, the guidance image / is filtered using a GF with both guidance image and image to be filtered being the same, i.e.

where

*l = A∑_fe _{e Wi} with b_k = (1 -a_k ). F_k, σ/(² is the variance of / in w_k, \w\ is the number of pixels in w_k, ε is a regularization parameter penalizing large a/r, and /_k = -^-∑i _{e Wfr} is the mean of / in

I I ^

[0086] Figure 3 shows (a) the raw guidance image /, and (b) the processed (de- noised) guidance image /', derived using the guidance image module 206 of the system of Fig. 2, for an example image.

Depth motion artifact suppression algorithm

[0087] Returning to Fig. 2, this shows the flow diagram of the method for motion artefact handling, in particular the use of motion artifact handling module 210, in an embodiment. With regard to processing at motion artifact handling module 210, a dedicated GF that integrates the plausibility map P, and using the guidance image /' is defined as follow:

where l^wfcl ∑] e w_k jj'D]P]-'kDk

- |w| ^{∑fe (}

|w| ∑fe ( and D_k = — "- Ei e w_t ^Di^pi is the mean of D in w_k weighted by the map P, \w\ is the constant number of pixels in the window w_t centered at pixel i, \w_k \ =

∑i e w_k Pi is the number of pixels in w_k where the plausibility is 1 , a_k = ^∑_{i e wfc} l?P - 11 where l_k = ^∑_{i £ Wfc} /^.

[0088] Figure 4 shows (a) the input depth map D, and (b) the processed (resulting) depth map D', derived using the system of Fig. 2, for the example image of Fig. 3. An improvement in image quality (e.g. sharpness, edge definition, noise reduction) is apparent. [0089] While embodiments have been described by reference to embodiments having various components in their respective implementations, it will be appreciated that other embodiments make use of other combinations and permutations of these and other components.

[0090] Furthermore, some of the embodiments are described herein as a method or combination of elements of a method that can be implemented by a processor of a computer system or by other means of carrying out the function. Thus, a processor with the necessary instructions for carrying out such a method or element of a method forms a means for carrying out the method or element of a method. Furthermore, an element described herein of an apparatus embodiment is an example of a means for carrying out the function performed by the element for the purpose of carrying out the invention.

[0091 ] In the description provided herein, numerous specific details are set forth. However, it is understood that embodiments of the invention may be practiced without these specific details. In other instances, well-known methods, structures and techniques have not been shown in detail in order not to obscure an understanding of this description.

[0092] Thus, while there has been described what are believed to be the preferred embodiments of the invention, those skilled in the art will recognize that other and further modifications may be made thereto without departing from the scope of the invention, and it is intended to claim all such changes and modifications as fall within the scope of the invention. For example, any formulas given above are merely representative of procedures that may be used. Functionality may be added or deleted from the block diagrams and operations may be interchanged among functional blocks. Steps may be added or deleted to methods described within the scope of the present invention.

Claims

Claims:

1 . A method for real-time motion artifact handling and noise removal for time-of- flight (ToF) sensor images, the ToF sensor images corresponding to received signals (r(t)) received by a ToF sensor following sending of modulated sent signals (s(t)), the method comprising:

calculating values of a cross correlation function C(T) at a plurality of temporally spaced positions or phases from the sent (s(t)) and received (r(t)) signals, thereby deriving a plurality of respective cross correlation values [C(T₀), c r^, C(T₂), C(T₃)] ;

deriving, from the plurality of cross correlation values [C(T₀), C(T₁), C(T₂), C(T₃)] , a depth map D, the depth map D having values representing, for each pixel, distance to a portion of an object upon which the sent signals (s(t)) are incident;

deriving, from the plurality of cross correlation values [C(T₀), C(T₁), C(T₂), C(T₃)] , a guidance image (/; /'), the guidance image (/; /') being an image having well defined edges; and

generating an output image D' based on the depth map D and the guidance image (/; /'), the output image D' comprising an edge-preserving and smoothed version of depth map D, the edge-preserving being from guidance image (/; /')■

2. The method of claim 1 , comprising acquiring a plurality of phase shifted images in succession, each phase shifted image corresponding to a respective temporally spaced position or phase.

3. The method of claim 1 or 2, wherein each cross correlation value [c(T₀), c(T_i), c(T₂), c(T-_i)] corresponds to one of a plurality of phase-shifted images, and the guidance image (/; /') is generated from the phase-shifted images.

4. The method of claim 2 or 3, wherein the phase-shifted images comprise amplitude images.

5. The method of any of the preceding claims, wherein deriving the guidance image comprises selecting as the guidance image a phase-shifted image from a plurality of previously-acquired phase-shifted images, based on one or more predetermined criteria.

6. The method of claim 5, wherein the predetermined criteria comprise that the phase-shifted image (i) with maximum amplitude of the object degraded by motion artefact, (ii) with maximum object edge sharpness value, (iii) with the best edge contrast and/or image SNR, (iv) with the maximum average spatial amplitude or (v) that is non-corrupted, is selected as the guidance image.

7. The method of any of the preceding claims, including using a guided filter (GF) to apply valid depth measurements to previously identified corrupted depth pixels due to motion artifacts.

8. The method of any of the preceding claims, wherein generating an output image D' comprises determining theoutput image D' as:

9. The method of any of the preceding claims, wherein generating an output image D' comprises:

filtering the guidance image / to generate a de-noised guidance image t , and generating an output image D' based on the depth map D and the de-noised guidance image /.

10. The method of claim 9, wherein filtering the guidance image / to generate a de-noised guidance image / comprises using a guided filter to perform said filtering.

1 1 . The method of any of the preceding claims, wherein generating an output image D' further comprises:

generating a plausibility map P based on the plurality of cross correlation values [c(T₀), c(T_i), c(T₂), c(T-_i)] , the plausibility map P comprising, for each pixel of the depth map D, a value indicative of whether the pixel is corrupted by motion or saturation; and generating the output image D' based on the depth map D and the plausibility map P, and on either the guidance image / or the de-noised guidance image /.

12. The method of claim 1 1 , wherein generating the plausibility map P comprises determining, for each pixel, a metric p_t representing the deviation of the phase-shifted amplitudes from a sinusoidal model of the cross-correlation function.

13. The method of claim 12, wherein the metric p_t comprises:

Pi = kOi) - C(T₀) - C(T₂) + c(r₃) \/(a + a)

where a is a regularization parameter preventing high value of p_t when the amplitude a is low.

14. The method of claim 12 or 13, wherein the plausibility map P has values P_t , for each pixel /^' such that:

p _ ( 1 if metric Pi > δ

1 <■ 0 otherwise

where δ is a motion threshold value, above which motion is detected at pixel /^'.

15. The method of claim 14, wherein δ is determined by capturing by the ToF sensor of an empty or motionless scene.

16. The method of claim 9, or any claim dependent thereon, wherein filtering the guidance image / to derive the de-noised guidance image /' comprises:

applying an edge preserving de-noising filter to guidance image /, whereby edge information and noise reduction from the guidance image / is transferred to the output image D'.

17. The method of claim 9, or any claim dependent thereon, wherein filtering the guidance image / comprises deriving de-noised guidance image /' using:

where

1 σ²

a, =— Y_k p _w . a_k, with a_k = -r^- with b_k = (1 — a_k 3k is the variance of / in w_k,

\w\ is the number of pixels in w_k,

ε is a regularization parameter penalizing large a,, and

18. The method of any of the preceding claims, when dependent on claims 9 and 1 1 , or any claim dependent thereon, wherein generating an output image D' comprises generating an output image D' according to:

Ό = a 't + b_v

where l^wfcl ∑] E w_k 'j'D]P]- Dk

~l |_w | ^K fc W; _σ2_+ε b_l = -^-_l∑_{k £ Wi} D_k-a_k l_k, and

D_k = — "- Ei e w_t

\w_k\ ^Di^pi is the mean of D in w_k weighted by the map P,

|w| is the constant number of pixels in the window w_t centered at pixel i, \^w _k\ = ∑i e w_fc Pi is the number of pixels in w_k where the plausibility is 1 , and

°l = Ι _Ρι .

19. The method of any of the preceding claims, wherein the output image D' comprise a version of depth map D alternatively or additionally from which motion artifacts and/or noise have been suppressed or removed.

20. The method of any of the preceding claims, wherein the cross correlation function c(r) is calculated as:

C(T) = j^__Q r(t). s(t + r)dt = h + ^ COS(O>T + φ), where the s(t) is the sent signal given by s(t) = 1 + cos(wt),

r(t) is the received signal given by r(t) = h + a. cos(wt - ψ)

ω = 2nf_m is the angular modulation frequency, f_m is the modulation frequency, and

h is the background light plus the non-modulated part of the sent signal.

21 . The method of claim 20, wherein the cross correlation values [C(T₀), C(T₁), C(T₂), C(T₃)] are determined from c(-r) as four samples r_k), k = 0, .. ,3 , taken at four subsequent time intervals r_k = k. - = k/4f_m within a modulated period T.

22. The method of claim 21 , wherein the distance measurements d for each pixel of the depth map D are obtained from

L

Ζπ

c = 3. 10⁸ m/s, the speed of light,

L is the working range or non-ambiguity distance range of the ToF sensor, and

23. The method of claim 4, or any claim dependent thereon, wherein an amplitude image A defined as A = [¾^■] _x . where the α_ί; are determined from: a = 1/2 V(C(T₃) - c(_Tl))² + (C(T₀) - C(T₂))² , where C{T_Q), C{T₁), C{T₂), C(T₃) are the cross correlation values.

24. The method of any of the preceding claims, wherein four cross correlation values [C(T₀), C(T₁), C(T₂), C(T₃)] are calculated from the cross correlation function c(r).

25. A programmable image processing system when suitably programmed for carrying out the method of any of the preceding claims, the system comprising circuitry for receiving or storing the received signals (r(t)) and sent signals (s(t)), and processing circuitry for performing the methods.

26. A system for real-time motion artifact handling and noise removal for time-of- flight (ToF) sensor images, the ToF sensor images corresponding to received signals (r(t)) received by a ToF sensor following sending of modulated sent signals (s(t)), the system comprising: circuitry for receiving or storing the received signals (r(t)) and sent signals processing circuitry, coupled to the circuitry for receiving or storing the received signals (r(t)) and sent signals (s(t)), the processing circuitry being operable for

calculating values of a cross correlation function C(T) at a plurality of temporally spaced positions or phases from the sent (s(t) ) and received (r(t)) signals, thereby deriving a plurality of respective cross correlation values [C(T₀), c r^, C(T₂), C(T₃)] ;

deriving, from the plurality of cross correlation values [C(T₀), C(T₁), C(T₂), C(T₃)] , a depth map D, the depth map D having values representing, for each pixel, distance to a portion of an object upon which the sent signals (s (t)) are incident;

generating an output image D' based on the depth map D and the guidance image (/; /'), the output image D' comprising an edge- preserving and smoothed version of depth map D, the edge-preserving being from guidance image (/; /')■

27. A recordable, rewritable or storable medium having recorded or stored thereon data defining or transformable into instructions for execution by processing circuitry and corresponding to at least the steps of any of claims 1 to 24.

28. A server computer incorporating a communications device and a memory device and being adapted for transmission on demand or otherwise of data defining or transformable into instructions for execution by processing circuitry and corresponding to at least the steps of any of claims 1 to 24.