Acoustic synchronous localization and mapping method based on single-cluster probability hypothesis density

Abstract

The invention discloses an acoustic synchronous localization and mapping method based on single-cluster probability hypothesis density. The method specifically comprises the following steps: carrying out recursive propagation on a sound source state and a weight value, a mean value and a covariance of probability hypothesis density observed by DoA; the distance minimum value and the distance maximum value from the sound source to the robot are uniformly distributed, a DoA observation value and sound source state missing distance information serve as the mean value of sound source state birth probability hypothesis density, and the mean value of the sound source state probability hypothesis density is optimized through a particle swarm optimization algorithm; updating the optimized sound source state probability hypothesis density mean value and covariance through an unscented Kalman filter; generating a robot pose particle weight by adopting a single cluster strategy, carrying out weighted average on the robot pose particle weight and Rao-Blackwelled filter particles to obtain robot pose estimation, trimming and combining the updated probability hypothesis density weight, the mean value and the covariance, and obtaining the robot pose estimation. And taking a probability hypothesis density mean value corresponding to the probability hypothesis density weight greater than a set threshold value as a sound source position, and taking the number of probability hypothesis density weights meeting the condition as the number of sound sources.

Description

Technical Field

[0001] The present invention relates to the field of acoustic synchronous positioning, and in particular to an acoustic synchronous positioning and mapping method based on single cluster probability hypothesis density. Background Art

[0002] With the continuous advancement of science and technology, robotics has made great progress in the past few decades. Robots have been widely used in industry, services, medical care, military and other fields, and have become an important part of modern society. SLAM (Simultaneous Localization and Mapping) technology is one of the core technologies in the fields of autonomous mobile robots, unmanned vehicles, etc., aiming to achieve autonomous positioning and map construction of robots in unknown environments. Acoustic SLAM (Acoustic SLAM) is a technology that uses acoustic information for simultaneous positioning and map construction. It aims to overcome the limitations of traditional SLAM in conditions such as changes in ambient light and poor light conditions, achieve stable positioning and map construction under a wider range of environmental conditions, and reduce dependence on environmental conditions. Acoustic sensors have unique advantages because sound, as a ubiquitous information carrier, can be obtained in most environments, and compared with other sensors, acoustic sensors are less sensitive to ambient light conditions and obstructions. Therefore, compared with traditional SLAM, acoustic SLAM is more robust.

[0003] In the prior art, such as in the paper "Acoustic SLAM With Moving Sound Event Based on Auxiliary Microphone Arrays", Hud et al. proposed an auxiliary microphone array acoustic SLAM (ASLAM) method based on dual interactive multiple models and unscented Kalman filter (D-IMM-UKF) for single moving sound source scenarios. First, the dual unscented Kalman filter (D-UKF) of ASLAM is derived by using the independent characteristics between the robot and the source, which can simultaneously locate and track the robot and the speaker. Compared with the traditional UKF, D-UKF has a more efficient data processing structure and avoids large-scale matrix inversion calculations. Then, an interacting multiple model is used for the different motion dynamics of the robot and the speaker in space. In order to avoid the underdetermination when there is only acoustic information, a small number of static microphone arrays are used as auxiliary information. Finally, the D-IMM-UKF algorithm is used to estimate the position of the mobile robot and the speaker. However, the disadvantage of this technology is that it requires setting up several groups of static auxiliary microphone arrays, which reduces its practicality in actual scenarios; this method also does not consider the influence of observation clutter. When clutter exists, the positioning accuracy is low and it is difficult to determine the number of sound sources.

[0004] In the paper "EchoSLAM: Simultaneous localization and mapping with acoustic echoes", Miranda et al. used acoustic echoes to perform simultaneous positioning and mapping in an unknown room, accurately estimate the geometry of the room, and track the trajectory of a robot moving in the room. The first algorithm is a triangulation-based positioning estimation algorithm. It assumes that the measurement results are ideal and noise-free, and the error of the robot motion is Gaussian distributed. The algorithm calculates the TOA (Time of Arrival) of the room impulse response (RIR) through three consecutive measurements to calculate the position of each wall. The second algorithm is a Bayesian solution inspired by FastSLAM. It takes into account the uncertainty of robot perception and motion. This method uses conditional independence to factorize the joint probability distribution of the robot position and wall parameters, thereby making efficient estimates through Bayesian filtering. However, the disadvantage of this technology is that using only one omnidirectional microphone and signal source may limit the accuracy or range of the measurement, which may affect the performance in complex environments or high noise or high interference. The TOA estimate obtained from the RIR may be inaccurate and susceptible to interference from other sound sources. In addition, this method ignores the impact of measurement clutter. Summary of the invention

[0005] In view of the problems existing in the prior art, the present invention discloses an acoustic synchronous positioning and mapping method based on single cluster probability hypothesis density, which specifically includes the following steps:

[0006] The sound source state and DoA observation data are modeled as random finite sets, and the probability hypothesis density weights, means and covariances of the sound source state and DoA observation data are recursively propagated;

[0007] The minimum and maximum distances from the sound source to the robot are evenly distributed, and the random numbers randomly generated by the even distribution are used as the missing distance information of the sound source state;

[0008] The DoA observation data and the missing distance information of the sound source state are used as the mean of the density of the probability hypothesis of the sound source state, and the particle swarm optimization algorithm is used to optimize the mean of the density of the probability hypothesis of the sound source state.

[0009] The mean and covariance of the optimized sound source state probability hypothesis density are updated through the unscented Kalman filter;

[0010] A single cluster strategy is used to generate the robot pose particle weights, and the robot pose particle weights are weighted averaged with the Rao-Blackwellized filter particles to obtain the robot pose estimation, and the robot pose particle weights are resampled.

[0011] The mean and covariance of the updated sound source state probability hypothesis density are trimmed and merged. When the weight of the probability hypothesis density is greater than the set threshold, the corresponding mean of the probability hypothesis density is the sound source position, and the number of weights of the probability hypothesis density that meets the conditions is the number of sound sources.

[0012] Furthermore, when the sound source state and DoA observation data are modeled as random finite sets: the number of constructions is N t The source state is a random finite set s t,n Indicates the status of each sound source

[0013]

[0014] B t Simulating the birth process at time t, P(s t-1,n ) represents the survival target between t-1 and t;

[0015]

[0016] The DoA estimation algorithm is used to infer the instantaneous direction of each sound source relative to the robot at time t. The multi-source measurement process is assumed to be the union of the measurement value generated by the target and the clutter caused by the reverberation. The measurement data Including quantity M t DoA observation data and robot measurement data The robot measurement data includes the robot running speed y t,v and direction y t,γ

[0017]

[0018] Among them, K t is a random set of false measurements generated by clutter, and each DoA observation data is Azimuth φ t,m ∈[0,2π) and the elevation angle θ t,m ∈[0,π).

[0019] Furthermore, the particle swarm optimization algorithm is used to optimize the mean of the probability hypothesis density of the sound source state, including: estimating the distance of each unmeasured sound source relative to the robot, and proposing J b The sound source state lacks distance information to initialize the mean value of the probability density of the i-th particle birth hypothesis

[0020]

[0021] Where U(r min ,r max ) represents the minimum distance r from the sound source to the robot min and the maximum value r max The mean of the birth probability density is assumed to be Built

[0022]

[0023] Among them, g -1 (·) is the transformation from spherical coordinates to Cartesian coordinates.

[0024] Furthermore, when the particle swarm optimization algorithm is used to optimize the mean of the probability hypothesis density of the sound source state: the Gaussian mixture probability hypothesis density is used to estimate the sound source position. The mean of the Gaussian term represents the possible location of the sound source. The sound source position estimation requires the prediction and update of the probability hypothesis density under the position and posture conditions of each robot particle. The prediction method of the probability hypothesis density is

[0025] λ tt-1 (s t,n |r 1:t ,Z 1:t-1 )=b(s t,n |r t )+λ t-1 (s t,n r 1:t-1 ,Z 1:t-1 ) (7)

[0026] Among them, b(s t,n |r t ) is the birth probability density, r t is the robot posture, is the prior probability hypothesis density of the i-th particle;

[0027] Update the mean of the predicted probability hypothesis density using the velocity and position formula in the particle swarm optimization algorithm

[0028]

[0029] Among them, rand1 and rand2 are standard normal distribution random numbers with positive diagonal elements, and p pbest and p gbest They represent the local optimal value and the global optimal value of the mean of the predicted probability hypothesis density, respectively, c1=c2, which represents the learning factor;

[0030] Introduce the fitness function to judge the optimization degree of the particle swarm algorithm for predicting the sound source

[0031]

[0032] Where R represents the noise covariance matrix, It is a prediction of the observed value of the sound source.

[0033] Furthermore, the mean and covariance of the optimized sound source state probability hypothesis density are updated through the unscented Kalman filter:

[0034] When the microphone array receives a sound source signal, the probability hypothesis density of each sound source is updated separately. The update formula of the probability hypothesis density is as follows

[0035]

[0036] Given the estimated robot pose The mean of the jth Gaussian component and covariance The unscented Kalman filter is used in the update process of formula (15), and the predicted state is given by the following formula Covariance matrix and the cross-covariance matrix Make an update

[0037]

[0038]

[0039] The updated probability hypothesis density mean and covariance are obtained using the following formulas (24) and (25);

[0040]

[0041] in, is the Kalman gain.

[0042] Furthermore, when the robot pose particle weights and the Rao-Blackwellized filter particles are weighted averaged to obtain the robot pose estimate:

[0043] Use the Rao-Blackwell particle filter to obtain an estimate of the robot's pose The robot's posture is represented by particles with different weights in

[0044]

[0045] The single cluster strategy is used to generate the robot pose particle weights. The single cluster strategy formula is as follows:

[0046]

[0047] According to resampling get

[0048] Due to the adoption of the above technical scheme, the present invention provides an acoustic synchronous positioning and mapping method based on single-cluster probability hypothesis density, which models the sound source state and orientation observation data as a random finite set, and recursively propagates its probability hypothesis density, and uses the uniform distribution of the minimum and maximum distances from the sound source to the robot as the missing distance information to construct a random range of the sound source state. In the probability hypothesis density prediction process, a fitness function is constructed using the latest observations through a particle swarm optimization algorithm for optimization; then, an unscented Kalman filter is used to calculate the distance hypothesis orientation information, and a single cluster strategy is used for the particle weighting process of the Rao-Blackwellized filter; finally, the number and position of the sound source and the motion trajectory of the robot are jointly estimated, and the proposed ASLAM shows commendable positioning accuracy even under voice inactive and clutter measurement conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0050] Figure 1 is a flow chart of the method of the present invention;

[0051] Figure 2 Schematic diagram of acoustic SLAM in the present invention;

[0052] Figure 3 This is a sample diagram of the acoustic SLAM results of the present invention. DETAILED DESCRIPTION

[0053] In order to make the technical solutions and advantages of the present invention more clear, the technical solutions in the embodiments of the present invention are clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention:

[0054] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0055] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0056] like Figure 1 The acoustic simultaneous positioning and mapping method based on single cluster probability hypothesis density shown in FIG. 1 specifically includes the following steps:

[0057] S1: Model the source state and DoA observation data as a random finite set. The number of constructions is N t The source state is a random finite set s t,n Represented as the state of each sound source.

[0058]

[0059] B t Simulating the birth process at time t, P(s t-1,n ) describes the survival goal between t-1 and t.

[0060]

[0061] The DoA estimation algorithm is used to infer the instantaneous direction of each sound source relative to the robot at time t. Assume that the multi-source measurement process is composed of the measurement value generated by the target and the clutter caused by reverberation. Measurement data By the quantity M t DoA observation data and robot measurement data The robot measurement data includes the robot running speed y t,v and direction y t,γ

[0062]

[0063] Among them, K t is a random set of false measurements generated by clutter, and each DoA observation is Azimuth φ t,m ∈[0,2π) and the elevation angle θ t,m ∈[0,π).

[0064] S2: Evenly distribute the minimum and maximum distances from the sound source to the robot, and use the random numbers randomly generated by the uniform distribution as the missing distance information of the sound source state. Estimate the distance of each unmeasured sound source relative to the robot by proposing J b Random range To initialize the mean value of the birth probability hypothesis density of the i-th particle

[0065]

[0066] Among them, U(r min ,r max ) represents the minimum distance r from the sound source to the robot min and the maximum value r max The mean of the birth probability density is assumed to be Built

[0067]

[0068] Among them, g -1 (·) is the transformation from spherical coordinates to Cartesian coordinates.

[0069] S3: The probability hypothesis density prediction process uses the particle swarm optimization algorithm to optimize the mean of the probability hypothesis density of the sound source state. The Gaussian mixture probability hypothesis density is used to estimate the sound source position, and the mean of the Gaussian term represents the possible location of the sound source. The sound source position estimation requires the prediction and update of the probability hypothesis density under the position and posture conditions of each robot particle. The prediction formula for the probability hypothesis density is

[0070] λ tt-1 (s t,n |r 1:t ,Z 1:t-1 )=b(s t,n |r t )+λ t-1 (s t,n r 1:t-1 ,Z 1:t-1 ) (7)

[0071] Among them, b(s t,n |rt ) is the birth probability density, r t is the robot posture, is the prior probability hypothesis density of the ith particle

[0072]

[0073] in, Expressed as the number of Gaussian mixture terms, and are the corresponding prediction weights, mean and covariance of the j-th Gaussian mixture term of the probability hypothesis density of the i-th particle, respectively.

[0074] The birth probability of particles is assumed to be a Gaussian mixture density.

[0075]

[0076] in, represents the number of Gaussian mixture terms assuming the birth probability density, and are the corresponding prediction weights, means and covariances of Gaussian mixture terms respectively. Therefore, the predicted probability hypothesis density can be written as follows.

[0077]

[0078] in, Expressed as the prior probability hypothesis density, it is and the birth probability density combination.

[0079] Update the mean of the predicted probability hypothesis density using the velocity and position formula in the particle swarm optimization algorithm

[0080]

[0081]

[0082] Among them, rand1 and rand2 are standard normal distribution random numbers with positive diagonal elements, and p pbest and p gbest They represent the local optimal value and the global optimal value of the predicted probability hypothesis density respectively, and c1=c2 is the learning factor.

[0083] Furthermore, a fitness function is introduced to determine the degree of optimization of the particle swarm algorithm for predicting sound sources.

[0084]

[0085] Where R represents the noise covariance matrix, It is a prediction of the observed value of the sound source.

[0086]

[0087] where h(·) is the observation model, g(·) is the transformation from Cartesian coordinates to spherical coordinates.

[0088] S4: The probability hypothesis density update process updates the weights, mean, and covariance of the optimized probability hypothesis density through the unscented Kalman filter. The specific method is: when the microphone array receives the sound source signal, the probability hypothesis density of each sound source is updated separately. The probability hypothesis density update formula is as follows

[0089]

[0090] Among them, P D The probability of detecting a feature.

[0091]

[0092]

[0093] Among them, k t =λ c VU is the probability hypothesis density of a random finite set of clutter, V is the room volume, λ c is the clutter rate, and the clutter obeys a uniform distribution.

[0094] The unscented Kalman function is applied to the update process of the mean and variance of the probability hypothesis density. The mean of the jth Gaussian component and variance Weighted sigma points It is composed of the following formula

[0095]

[0096] Where ε = α 2 (n+κ)-n, α=0.01 is the scale parameter, κ=1, β=2. The predicted measurement is generated by

[0097]

[0098] The predicted state is given by the following formula Covariance matrix and the cross-covariance matrix renew.

[0099]

[0100] The updated probability hypothesis density mean and covariance can be obtained through formulas (18) to (23).

[0101]

[0102] in, is the Kalman gain.

[0103] S5: The robot pose estimation is obtained by weighted averaging the robot pose particle weights and the Rao-Blackwellized filter particles. The specific method is: In order to obtain the robot pose estimation, the Rao-Blackwell particle filter is used to estimate The robot's posture is represented by particles with different weights in

[0104]

[0105] The single cluster strategy is used to generate the robot pose particle weights. The single cluster strategy formula is as follows:

[0106]

[0107] Among them, resampling Can get

[0108] S6: The sound source position is estimated by trimming and merging the updated optimized probability hypothesis density weights, mean and covariance, and the mean of the probability hypothesis density corresponding to the probability hypothesis density weight greater than the set threshold is the sound source position, and the number of probability hypothesis density weights that meet the conditions is the number of sound sources.

[0109] In order to verify the effectiveness of the present invention, the Imgae model was used to simulate a 6×6×3m 3 In the room, two static sound sources are located at s1 = (1, 3, 1.7) m and s2 = (5, 4, 1.75) m. The mobile robot is equipped with a microphone array and an IMU (Inertial Measurement Unit) to obtain DoA observation data and the robot speed y t,v and direction y t,γ To ensure that the mobile robot is within the room, it is necessary to keep the robot within 1m of the room boundary. to enforce the maximum turn. The sound source is a pure speech signal with a sampling frequency of 16kHz randomly selected from the TIMIT database. DoA is calculated using the classic phase-transformed generalized cross-correlation algorithm (GCC-PHAT). RMSE (root mean square error), MAE (mean absolute error) and OSPA (ptimal Subpattern Assignment distance) are used to evaluate the performance of the acoustic synchronous localization and mapping method. Under different signal-to-noise ratios (SNR) and different reverberation conditions, the positioning results of 50 Monte Carlo experiments of the acoustic synchronous localization and mapping method are shown in Tables 1 and 2. It can be seen from the test results that under low signal-to-noise ratio conditions, the impact on the positioning results is still small, which can meet the requirements of practical applications. The impact on the method in a reverberant environment is small, which can meet the requirements of practical applications. All of these show the good robustness and good working stability of the present invention. Figure 3 Table 3 and Table 3 show the positioning results of different robot trajectories and sound source positions when the reverberation time is 300ms and the signal-to-noise ratio is 25dB.

[0110] Table 1 Positioning error of acoustic synchronous localization and mapping method under different SNRs with 300ms reverberation time

[0111]

[0112] Table 2 Positioning error of acoustic synchronous localization and mapping method under different reverberation times at 25dB SNR

[0113]

[0114]

[0115] Table 3 Comparison of different indicators of four groups of trajectories

[0116]

[0117] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. An acoustic simultaneous localization and mapping method based on single cluster probability hypothesis density, characterized in that include: The sound source state and DoA observation data are modeled as random finite sets, and the probability hypothesis density weights, means and covariances of the sound source state and DoA observation data are recursively propagated; The minimum and maximum distances from the sound source to the robot are evenly distributed, and the random numbers randomly generated by the even distribution are used as the missing distance information of the sound source state; The DoA observation data and the missing distance information of the sound source state are used as the mean of the density of the probability hypothesis of the sound source state, and the particle swarm optimization algorithm is used to optimize the mean of the density of the probability hypothesis of the sound source state. The mean and covariance of the optimized sound source state probability hypothesis density are updated through the unscented Kalman filter; A single cluster strategy is used to generate the robot pose particle weights, and the robot pose particle weights are weighted averaged with the Rao-Blackwellized filter particles to obtain the robot pose estimation, and the robot pose particle weights are resampled. The mean and covariance of the updated sound source state probability hypothesis density are trimmed and merged. When the weight of the probability hypothesis density is greater than the set threshold, the corresponding mean of the probability hypothesis density is the sound source position, and the number of weights of the probability hypothesis density that meets the conditions is the number of sound sources.

2. The method for acoustic simultaneous localization and mapping based on single cluster probability hypothesis density according to claim 1, characterized in that: When the sound source state and DoA observation data are modeled as random finite sets: the number of constructions is N t The source state is a random finite set s t,n Indicates the status of each sound source B t Simulating the birth process at time t, P(s t-1,n ) represents the survival target between t-1 and t; The DoA estimation algorithm is used to infer the instantaneous direction of each sound source relative to the robot at time t. The multi-source measurement process is assumed to be the union of the measurement value generated by the target and the clutter caused by the reverberation. The measurement data Including quantity M t DoA observation data and robot measurement data The robot measurement data includes the robot running speed y t,v and direction y t,γ Among them, K t is a random set of false measurements generated by clutter, and each DoA observation data is Azimuth φ t,m ∈[0,2π) and the elevation angle θ t,m ∈[0,π).

3. The method for acoustic simultaneous localization and mapping based on single cluster probability hypothesis density according to claim 2, characterized in that: The particle swarm optimization algorithm is used to optimize the mean of the probability hypothesis density of the sound source state, including: estimating the distance of each unmeasured sound source relative to the robot, and proposing J b The sound source state lacks distance information to initialize the mean value of the probability density of the i-th particle birth hypothesis in Represents the minimum distance r from the sound source to the robot min and the maximum value r max The mean of the birth probability density is assumed to be Built Among them, g -1 (·) is the transformation from spherical coordinates to Cartesian coordinates.

4. The method for acoustic simultaneous localization and mapping based on single cluster probability hypothesis density according to claim 3, characterized in that: When using the particle swarm optimization algorithm to optimize the mean of the probability hypothesis density of the sound source state: the Gaussian mixture probability hypothesis density is used to estimate the sound source position. The mean of the Gaussian term represents the possible location of the sound source. The sound source position estimation requires the prediction and update of the probability hypothesis density under the position and posture conditions of each robot particle. The prediction method of the probability hypothesis density is λ tt-1 (s t,n |r 1:t ,Z 1:t-1 )=b(s t,n |r t )+λ t-1 (s t,n r 1:t-1 ,Z 1:t-1 ) (7) Among them, b(s t,n |r t ) is the birth probability density, r t is the robot posture, is the prior probability hypothesis density of the i-th particle; Update the mean of the predicted probability hypothesis density using the velocity and position formula in the particle swarm optimization algorithm Among them, rand1 and rand2 are standard normal distribution random numbers with positive diagonal elements, and p pbest and p gbest They represent the local optimal value and the global optimal value of the mean of the predicted probability hypothesis density, respectively, c1=c2, which represents the learning factor; Introduce the fitness function to judge the optimization degree of the particle swarm algorithm for predicting the sound source Where R represents the noise covariance matrix, It is a prediction of the observed value of the sound source.

5. The method for acoustic simultaneous localization and mapping based on single cluster probability hypothesis density according to claim 4, characterized in that: The mean and covariance of the probability hypothesis density of the optimized sound source state are updated through the unscented Kalman filter: When the microphone array receives a sound source signal, the probability hypothesis density of each sound source is updated separately. The update formula of the probability hypothesis density is as follows Given the estimated robot pose The mean of the jth Gaussian component and covariance The unscented Kalman filter is used in the update process of formula (15), and the predicted state is given by the following formula Covariance matrix and the cross-covariance matrix Make an update The updated probability hypothesis density mean and covariance are obtained using the following formulas (24) and (25); in, is the Kalman gain.

6. The method for acoustic simultaneous localization and mapping based on single cluster probability hypothesis density according to claim 5, characterized in that: When the robot pose estimation is obtained by weighted averaging the robot pose particle weights and the Rao-Blackwellized filter particles: Use the Rao-Blackwell particle filter to obtain an estimate of the robot's pose The robot's posture is represented by particles with different weights in The single cluster strategy is used to generate the robot pose particle weights. The single cluster strategy formula is as follows: According to resampling get

Images