CN115469576A - A Teleoperation System Based on Hybrid Mapping of Human-Robot Arm Heterogeneous Motion Space - Google Patents

Abstract

The invention discloses a manipulator teleoperating system based on hybrid mapping of human-manipulator heterogeneous motion space, which uses a camera to identify the joint position information of the operator and performs data processing to obtain part of the joint angles of the operator and the coordinate pose of the end of the arm , map the pose of the end of the human arm to the end of the manipulator, and solve the inverse kinematics of the manipulator to obtain the expected joint angle of the manipulator. For the situation where there are multiple solutions or no solution exists, adjust the joint angle of the operator with reference to the joint angle of the operator, and according to The joint angle is expected to plan the motion of the robotic arm, control the motion of the real robotic arm, and make the robotic arm mimic the posture of the human arm as much as possible.

Description

A teleoperation system based on hybrid mapping of human-robot heterogeneous motion space

technical field

The invention belongs to the field of human-computer interaction, and relates to a method for remotely controlling a robot based on optical attitude capture and attitude mapping.

Background technique

Limited by the current level of robot intelligence, the development of a fully autonomous robot is still an unattainable goal for a long time, especially in dangerous, long-distance or unstructured environments such as nuclear radiation, space, and deep sea. Organically combining human's advanced decision-making and flexible operation capabilities with robots, and developing human-in-the-loop fine teleoperation technology will be an inevitable choice for a long time in the future.

Although the field of teleoperated robots has made great progress, and some of them have even been commercialized and put into practical use, the way operators command slave robots has not changed much. Generally speaking, the operator directly sends specific instructions to the slave robot while visually monitoring the behavior of the slave robot in the remote environment, which is called direct teleoperation, which often brings heavy cognitive burden to the operator. burden. In addition, the delay and bandwidth limitation of the communication between the operator and the remote robot have been the main factors restricting the performance of the teleoperation system; when using the traditional teleoperation system, the operator's two important perceptions (vision and force) sense) is deprived, seriously affecting the operator's ability to complete teleoperation tasks. Especially in the face of unstructured environments and unexpected accidents, human operators cannot deal with problems in time, which will reduce the success rate of tasks, and even bring hidden dangers to users' personal safety, and will also make operators in a state of high tension and prone to accidents. Mental fatigue reduces work efficiency.

In order to solve the above problems, researchers have conducted a lot of research, which can be roughly classified into two categories: First, combining the special performance of human-computer interaction technology based on visual, tactile, muscle nerve and other sensory signals can improve the work of teleoperated robot systems. Efficiency and control accuracy; Second, research and develop man-machine remote operation control mode in a targeted manner, give full play to the respective advantages of human operators and robots, improve the sensitivity and accuracy of the system while alleviating the fatigue of human operators, and realize man-machine Tasks are properly configured.

To teleoperate a robot, motion commands need to be generated, and optical-based pose capture is the least physically constrained of all solutions. Commercial marker-based products, such as OptiTrack, aim to capture poses by tracking reflective markers [Spanlang B, Navarro X, Normand J M, Kishore S, Pizarro R, Slater M. Real time whole body motion mapping for avatars and robots [A] .In:the 19thACM Symposium on Virtual Reality Software and Technology[C].Singapore:Association for Computing Machinery,2013:175-178.]. However, this approach consists of a time-consuming tagging process for operators to use all tags correctly, which makes it less widely applicable. For example, the markerless motion capture system Capture Live promises better convenience, but the high price of equipment and licenses has also greatly limited their popularity. In addition, these commercial methods all require operators to operate in a limited indoor workspace surrounded by camera arrays, resulting in less environmental flexibility. Today, portable depth detection devices such as the famous Microsoft Kinect, Intel RealSense, and LeapMotion are becoming more and more feasible. Researchers have made it possible to reduce the number of motion capture cameras (or even just one) and greatly simplify the motion capture process. .

Human-robot pose mapping, i.e. converting human poses to robot poses, bridges the last gap between human-robot motion. Luo et al [Luo R C, Shih B H, Lin T W.Real time human motionimitation of anthropomorphic dual arm robot based on Cartesian impedance control[A].In:IEEE International Symposium on Robotic and SensorsEnvironments[C].Piscataway:IEEE,2013: 25-30.] Online teaching of a 6-DOF manipulator is realized using Cartesian spatial impedance control that avoids the inverse kinematics calculation process. However, the motion mapping in Cartesian space only focuses on the synchronization of the end pose of the man-machine, and the motion trajectories of the shoulder and elbow joints of the robotic arm need to be planned manually. Nguyen et al. [NguyenV V, Lee J H, et al.Full-body imitation of human motions with Kinect and heterogeneous kinematic structure of humanoid robot[A].In:IEEE/SICEInternational Symposium on System Integration[C].Piscataway:IEEE,2012 :93-98.] First establish the joint space mapping relationship between the human arm and the 3-DOF robotic arm, and then use the space vector method to solve the joint angle of the human arm and pass it to the robotic arm as a control command to realize real-time motion simulation. Wu Weiguo et al [Wu Weiguo, Li Hua, Gao Liyang. Biped robot following walking and experiment captured by human walking[J]. Journal of Harbin Institute of Technology, 2017,49(1):21-29.] Performed six-freedom human lower limbs After the degree inverse kinematics calculation, the walking samples are established, and then the motion similarity index is used for sample splicing to drive the robot to achieve walking follow.

In the actual vision-based teleoperation process, due to camera noise, limitations in human joint recognition, and errors in human bone recognition, it is difficult to ensure that non-anthropomorphic robots can well imitate human arm poses using a single mapping method. The end pose mapping can make the end of the robot follow the human wrist well, but there are many inverse kinematics solutions and no solutions. Joint angle mapping can enable the robot to better imitate the posture of the human arm, but the number of joints in the human arm is not equal to the number of joints in the robot, and some joint angles of the human arm are difficult to measure.

Contents of the invention

In view of the above problems, the present invention proposes a teleoperation system based on the hybrid mapping of human-robot heterogeneous motion space, which can identify the joints of the human body and map them to the robotic arm, which reduces the restrictions on human motion and enables the robotic arm to do A gesture similar to that of a human arm has been achieved, and better results have been achieved in teleoperation.

The present invention is based on the teleoperation system of human-manipulator arm heterogeneous motion space hybrid mapping, which is divided into two modules, namely, the human-manipulator arm motion mapping module and the manipulator arm motion planning and control system.

Among them, the human-robot motion mapping module is constructed, and the human joint angle, position and posture recognition system is built. The human joint position recognition system of the Kinect camera is used to obtain the human joint position data and draw the human skeleton, and the Kalman filter is performed on the human joint position data, and then Calculate and output the pose and joint angle of the end of the human arm.

And using the human-machine motion space hybrid mapping method, first map the end pose of the human arm to the end of the robotic arm, and then solve the inverse kinematics of the robotic arm; further add joint mapping to obtain the optimal solution for the joint angle of the robotic arm.

The robot arm motion planning and control system uses socket to establish the communication transmission between Windows and ROS, publishes the expected joint angle data to the ROS node, and uses Moveit to subscribe and perform motion planning, and finally transmits the control information to the real robot arm to control the real machine arm movement.

Advantages of the present invention:

1. The present invention is based on the teleoperation system based on the hybrid mapping of human-robot heterogeneous motion space, which adopts unmarked motion recognition, which has less movement constraints on the operator, and only needs a Kinect camera to collect human joint data, and the hardware configuration requirements Low.

2. The present invention is based on the teleoperation system of human-manipulator arm heterogeneous motion space hybrid mapping, which solves the problem of multiple solutions and non-existence of the inverse kinematics of the manipulator caused by the terminal attitude mapping.

3. The present invention is based on the teleoperation system of human-manipulator arm heterogeneous motion space hybrid mapping. Only according to the Kinect camera to collect human joint data, it is impossible to calculate all the joint angles of the human arm. The hybrid mapping method is used to avoid this problem. Three joint angles need to be measured.

4. The present invention is based on the teleoperation system based on the hybrid mapping of human-robot heterogeneous motion space. The robot arm can better imitate the posture of the human arm and make actions similar to the posture of the human arm. It is easier for the operator to control it to complete the operation Task.

Description of drawings

Fig. 1 is the overall block diagram of the teleoperation system based on the mixed mapping of human-robot heterogeneous motion space in the present invention;

Fig. 2 is the kinematics model of human arm;

Figure 3 is the kinematics model of the ur5 robotic arm;

Fig. 4 is a parameter conversion diagram between connecting rods;

Figure 5 shows 25 three-dimensional joint positions relative to the camera coordinates;

Figure 6 is a human right arm model.

detailed description

The present invention will be described in further detail below in conjunction with the accompanying drawings.

The present invention is based on the teleoperation system of human-robot heterogeneous motion space hybrid mapping, as shown in Figure 1, which is divided into two modules, namely, the human-robot motion mapping module and the robot motion planning.

Among them, the human-robot motion mapping module uses the Kinect camera to obtain human joint information, and draws the human skeleton; uses any six-degree-of-freedom robotic arm as the object of teleoperation control, and maps the pose of the end of the human arm and the angles of each joint of the human arm The end of the manipulator and each joint of the manipulator are mapped separately, and the mappings of the two are combined to form a hybrid mapping method, and finally the desired joint angle of the manipulator is obtained.

The motion planning and control module of the manipulator is realized by the ROS robot control system. According to the expected joint angle of the manipulator, it is planned in Moveit, simulated in Gazebo, and controls the real manipulator movement.

The kinematics model is established in the human-robot motion mapping module. The kinematics model is the basis for mapping the pose and joint angle of the end of the arm. The specific establishment method adopts the D-H method. On the basis of the kinematic modeling of the robot, 5 Two kinematic models of -DOF arm model and UR5 robot model; among them, the 5-DOF arm model includes shoulder 3-DOF and elbow 2-DOF; Ur5 robot model is a 6-DOF mechanical arm; it can be seen that the two models are Heterogeneous models, as shown in Figure 2 and Figure 3.

The definition of DH parameters is shown in Figure 4. In Figure 4, z _i and z _i-1 coincide with the rotation axes of the i-th and i-1 joints of the manipulator respectively, and x _i-1 is along the i-1 joint axis and the i-th joint axis. The direction of the common vertical line a _i-1 of the joint axis; x _i is along the i-th joint axis and the direction of the common vertical line a _i of the i-th joint axis; α _i-1 is the distance between z _i-1 and z _i Included angle, a _i-1 is the connection length from z _i-1 to z _i measured along x _i-1 , where x _i-1 is perpendicular to z _i-1 and z _i ; d _i is x _i-1 to z i The link offset of _xi measured along _zi ; θi is the angle between _{xi -1} and _xi .

According to the D-H method, the coordinate transformation description from the i-1 joint coordinate system of the manipulator to the i joint coordinate system is shown in formula (1).

Further, the coordinate transformation matrix of the human arm from the basic coordinate system to the terminal coordinate system can be derived from formula (1)

Similarly, the transformation matrix of the 6-DOF manipulator in formula (3) is obtained as

In the above formula, [n _x , n _y , n _z ; o _x , o _y , o _z ; a _x , a _y , a _z ] and [p _x p _y p _z ] denote the posture and position of the end effector, respectively .

The Kinect camera in the described human-mechanical arm motion mapping model acquires the human body joint data and performs filtering processing as follows:

Human body joint data is obtained directly through Kinect, which contains 25 three-dimensional joint positions relative to the camera coordinates, as shown in Figure 5, including: spine foundation, mid-spine, neck, head, right shoulder, right elbow, right wrist, right Palm, left shoulder, left elbow, left wrist, left palm, right hip, right knee, right ankle, right foot, left hip, left knee, left ankle, left foot, intersection of spine and shoulder, right tip, right thumb, left hand Pointer, left thumb.

When the human body remains still, the position of the joint points collected by Kinect is not constant, but fluctuates around the actual value; there are two reasons for this, one is that the static human body is not really static; the other is that Kinect and body tracking methods collect data has errors. At the same time, the coordinates of the raised joint points will affect the description of the coordinates of the human wrist. In order to improve the accuracy of the system, it is necessary to filter the collected data through Kalman filtering before using the data, as follows:

x _k ＝A _k x _k-1 +B _k u _k +w _k z _k ＝H _k x _k +v _k (4)

x _k is the state vector, u _k is the state control vector, z _k is the measurement vector. A _k is the state transition matrix, B _k is the control input matrix, and H _k is the state observation matrix. w _k represents the process noise and v _k represents the measurement noise. w _k and v _k obey Gaussian distribution, and their covariance matrices are Q and R respectively, that is, w _k ~N(0,Q), v _k ~N(0,R).

In the human-mechanical arm motion mapping module, after collecting the data of human joints, the pose and joint angle of the end of the human arm are calculated, and the method is as follows:

手腕坐标系统

和Kinect坐标

需要左肩、右肩和脊柱的平面来构建坐标系，图中A点为左肩，G点为脊柱中段，O点为右肩，C点为右手腕，D点、E点、F点分别为右拇指、右手与右手指尖；上述各点均为kinect相机直接测量获得，与相机识别的25个点相对应。When the human body is not perpendicular to the ground, the basic coordinates of the human arm will be wrong. In order to solve this problem, the geometric model of the right arm of the human body is established in the present invention, as shown in FIG. 5 . The model has three coordinate systems, namely the shoulder coordinate system

Wrist Coordinate System

and Kinect coordinates

The planes of the left shoulder, right shoulder and spine are needed to construct the coordinate system. In the figure, point A is the left shoulder, point G is the middle of the spine, point O is the right shoulder, point C is the right wrist, and points D, E, and F are the right Thumb, right hand and right fingertips; the above points are directly measured by the kinect camera, corresponding to the 25 points recognized by the camera.

Let the above point O be the origin of the basic coordinates, and according to the geometric model of the right arm of the human body established above, then:

① The calculation method of the end pose of the human arm is:

与向量

方向相同，

与平面AGO的法向量相同。因此，

是由右手坐标系的规定得到的，如下：

with vector

same direction,

Same as normal vector for planar AGO. therefore,

It is obtained by the regulation of the right-handed coordinate system, as follows:

Point C is the origin of the coordinates of the end effector (ie, the right wrist). Therefore, the position of the wrist joint is:

表示平面DEF的法向量，

与向量

方向相同。

和

都是由右手得到的，如下：use

represents the normal vector of the plane DEF,

with vector

same direction.

with

are obtained by the right hand, as follows:

② Calculation method of human arm joint angle

在YOZ平面上的投影与

的夹角获得的，大臂与小臂的夹角θ₂是向量

与

的夹角，手腕的夹角θ₃用

与

的夹角表示。As shown in Figure 6, the angle θ ₁ between the boom and the YOZ plane is obtained by calculating the vector

The projection on the YOZ plane is the same as

The included angle is obtained, and the included angle θ ₂ between the arm and the forearm is a vector

and

The included angle, the included angle of the wrist θ ₃ is used

and

angle representation.

After obtaining the pose and joint angle of the end of the human arm through the above, the pose of the end of the human arm and the angles of each joint of the human arm can be mapped to the end of the manipulator and each joint of the manipulator. The specific mapping process is as follows:

Ⅰ. Perform terminal attitude mapping

The point-to-point method is used to map the position of the end of the human arm to the position of the end effector of the robot. Combining the two transformation matrices obtained by formula (2) and formula (3), the angles of each joint of the manipulator are obtained by solving the inverse kinematics. The mapping formula is as follows:

where [x _r , y _r , z _r ] ^T is the robot end effector position; [x _h , y _h , z _h ] ^T is the human arm end effector position; k _x , _ky , k _z are the ratio Coefficients; l _x , _ly , l _z are translational distances of the coordinates; k _x , _ky , k _z , l _x , _ly , l _z are calculated according to the actual size of the human arm and the mechanical arm.

In general, there will be multiple solutions for the inverse kinematics solution of the manipulator. The present invention proposes that the optimal solution should satisfy the following requirements for optimal solution selection:

1) In order to ensure the continuity and stability of the movement of the manipulator, the angle value of the first joint at the current moment (that is, the rotation angle of the manipulator base) is selected to have the smallest difference from the previous moment;

2) In order to make the robot arm make a posture similar to that of the human arm, on the basis of ₁ ), select the second joint angle value with the smallest difference from θ1. According to the above method, a group of optimal solutions can be finally determined.

Ⅱ. Add joint angle mapping

Due to the limited recognition accuracy of the Kinect camera, there are two problems: 1) the inverse kinematics solution does not exist; 2) there is a certain gap between the pose of the robot arm and the expected pose.

In the case that the inverse kinematics solution does not exist, the angle values of the 2nd to 4th joints of the manipulator cannot be solved, and the joint angles θ ₁ , θ ₂ , θ ₃ of the human arm can be directly mapped to the 2nd to 4th joint angles of the manipulator. joint.

Aiming at the problem of inconsistency in attitude, joint angle mapping is added, and the second and third joint angle values in the optimal solution obtained above are changed to θ ₁ and θ ₂ .

The control effect after adding the joint angle mapping is significantly better than the simple end pose mapping, thus realizing the hybrid mapping of the heterogeneous motion space of the human-robot arm.

Described manipulator motion planning and control module adopts socket to realize the communication of Windows end and Linux end ROS system, publishes the final expected joint angle value obtained by the aforementioned method to the ROS node, subscribes topic message at ROS end, in order to reduce time Delay, subscribe to the latest news on the topic node every time. Moveit is a function package of a series of mobile operations in ROS, mainly including motion planning, collision detection, kinematics, 3D perception, operation control and other functions, which can establish communication with real robotic arms. Therefore, Moveit is used to plan the forward kinematics trajectory of the manipulator, and finally realize the control of the real manipulator.

Based on the above-mentioned teleoperation system, the present invention proposes the specific flow of teleoperation control as follows:

A, obtain the human body joint coordinates collected by the Kinect camera;

B. Filter the human joint data, and calculate the pose and joint angle of the end of the human arm;

C. Calculate the mapping parameters k _x , _ky , k _z , l _x , _ly , l _z according to the real length of the human arm and the robotic arm, and perform terminal pose mapping;

D. Construct an inverse kinematics solution function according to the D-H parameters of the mechanical arm;

E. Select the optimal solution according to the optimal solution selection principle;

F. Add joint angle mapping;

G. Realize communication and publish the desired joint angle to the ROS node;

H. Subscribe and process topic messages, use Moveit for motion planning, and simulate in Gazebo;

G. The operator is making an action on the Kinect camera, and the robotic arm follows the movement of the operator's right arm to make a corresponding gesture, and interacts in real time to complete the teleoperation task.

Claims (7)

1. A teleoperation system based on human-mechanical arm heterogeneous motion space hybrid mapping is divided into two modules, namely a human-mechanical arm motion mapping module and a mechanical arm motion planning and control system; the method is characterized in that:

the human-mechanical arm motion mapping module is constructed, a human joint angle and position posture recognition system is built, human joint position data are obtained by adopting the human joint recognition system of the Kinect camera, a human skeleton is drawn, kalman filtering is carried out on the human joint position data, and then the tail end pose and the joint angle of the human arm are calculated and output;

a human-machine motion space mixed mapping method is adopted, the pose of the tail end of the human arm is mapped to the tail end of the mechanical arm, and inverse kinematics solution is carried out on the mechanical arm; further adding joint mapping to obtain an optimal solution of the joint angle of the mechanical arm;

the mechanical arm motion planning and control system establishes communication transmission between Windows and the ROS by using a socket, issues expected joint angle data to ROS nodes, uses Moveit subscription to perform motion planning, finally transmits control information to a real mechanical arm, and controls the real mechanical arm to move.

2. The teleoperation system based on the human-mechanical arm heterogeneous motion space hybrid mapping, as claimed in claim 1, wherein: establishing a kinematics model in a human-mechanical arm movement mapping module, wherein the establishment method adopts a D-H method, and two kinematics models, namely a 5-DOF arm model and a UR5 robot model, are established on the basis of robot kinematics modeling; wherein the 5-DOF arm model includes a shoulder 3-DOF and an elbow 2-DOF; the Ur5 robot model is a 6-DOF mechanical arm.

3. The teleoperation system based on the human-mechanical arm heterogeneous motion space hybrid mapping, as claimed in claim 1, wherein: the Kalman filtering method for the human body joint position data comprises the following steps:

x _k ＝A _k x _k-1 +B _k u _k +w _k z _k ＝H _k x _k +v _k

x _k is a state vector, u _k Is a state control vector, z _k Is a measurement vector; a. The _k Is a state transition matrix, B _k Is a control input matrix, H _k Is a state observation matrix; w is a _k Representative of process noise, v _k Representative of measurement noise; w is a _k And v _k Obeying a Gaussian distribution with covariance matrices Q and R, respectively, i.e. w _k ～N(0,Q)，v _k ～N(0,R)。

4. The teleoperation system based on the human-mechanical arm heterogeneous motion space hybrid mapping, as claimed in claim 1, wherein: the method for calculating the end pose and the joint angle of the human arm comprises the following steps:

a geometric model of the right arm of the human body is established, and the geometric model is provided with three coordinate systems which are respectively a shoulder coordinate system

Wrist coordinate system

And Kinect coordinates

Let point A be the left shoulder, point G be the middle section of the spine, point O be the right shoulder, point C be the right wrist, point D, point E, point F be the right thumb, right hand and right finger tip respectively;

according to the geometric model of the right arm of the human body, which is established as above, the following steps are carried out:

(1) the method for calculating the pose of the tail end of the human arm comprises the following steps:

and vector

The direction is the same as that of the first and second guide rails,

the normal vector is the same as the planar AGO; therefore, the temperature of the molten metal is controlled,

is derived from the specification of the right hand coordinate system as follows:

thus, the positions of the wrist joints are:

by using

A normal vector of the plane DEF is represented,

and vector

The directions are the same;

and

all obtained from the right hand, as follows:

(2) human arm joint angle calculation method

Angle theta between the large arm and the YOZ plane ₁ By calculating vectors

Projection on YOZ plane

Obtaining the included angle of the angle; the included angle theta between the big arm and the small arm ₂ Is a vector

And

angle of the wrist theta ₃ By using

And with

Is shown.

5. The teleoperation system based on the human-mechanical arm heterogeneous motion space hybrid mapping, as claimed in claim 1, wherein: the man-machine motion space hybrid mapping method comprises the following steps:

mapping the tail end position of the arm of the human hand to the position of an end effector of the robot by adopting a point-to-point method, combining a coordinate transformation matrix of the arm of the human body from a basic coordinate system to a terminal coordinate system and a transformation matrix of the arm with 6 degrees of freedom, and solving by inverse kinematics to obtain each joint angle of the arm; the mapping formula is as follows:

wherein, [ x ] _r ,y _r ,z _r ] ^T Is the position of the robot end effector; [ x ] of _h ,y _h ,z _h ] ^T Is the human arm end effector position; k is a radical of formula _x 、k _y 、k _z Is a proportionality coefficient; l _x 、l _y 、l _z Is the translation distance of the coordinates; k is a radical of _x 、k _y 、k _z 、l _x 、l _y 、l _z Calculating according to the actual sizes of the human arm and the mechanical arm;

the optimal solution should satisfy:

1) In order to ensure the continuity and stability of the motion of the mechanical arm, the difference value between the value of the 1 st joint angle at the current moment and the previous moment is the minimum;

2) In order to make the mechanical arm to make a similar posture as the human arm, on the basis of 1), theta is selected ₁ The 2 nd joint angle value with the smallest difference; finally, a set of optimal solutions is determined.

6. The teleoperation system based on the human-mechanical arm heterogeneous motion space hybrid mapping, as claimed in claim 1, wherein: the joint mapping method comprises the following steps:

aiming at the condition that the inverse kinematics solution does not exist, the 2 nd to 4 th joint angle values of the mechanical arm cannot be solved, so that the joint angle theta of the human arm is determined ₁ 、θ ₂ 、θ ₃ Directly mapping to the 2 nd to 4 th joints of the mechanical arm;

aiming at the problem of inconsistent postures, joint angle mapping is added, and the 2 nd and 3 rd joint angle values in the obtained optimal solution are changed into theta ₁ 、θ ₂ 。

7. The teleoperation system based on human-mechanical arm heterogeneous motion space hybrid mapping of claim 1, wherein: the specific process of teleoperation control is as follows:

A. acquiring human body joint coordinates acquired by a Kinect camera;

B. filtering the human joint data, and calculating the terminal pose and the joint angle of the human arm;

C. calculating mapping parameters according to the real lengths of the human arm and the mechanical arm, and mapping the end pose;

D. constructing an inverse kinematics solving function according to the D-H parameters of the mechanical arm;

E. selecting an optimal solution according to an optimal solution selection principle;

F. adding joint angle mapping;

G. communication is achieved, and the expected joint angle is issued to the ROS node;

H. subscribing and processing topic messages, performing motion planning by using Moveit, and performing simulation in a Gazebo;

G. and (5) finishing the teleoperation task in real time in an interactive manner.

Images