Method for training and operating movement estimation of objects
Learning extraction of movement information from sensor data includes providing a time series of frames of sensor data recorded by physical observation of an object, providing a time series of object boundary boxes each encompassing the object in sensor data frames, supplying the object boundary box at a time t, as well as a history of sensor data from the sensor data time series, and/or a history of object boundary boxes from the time series of object boundary boxes, prior to time t to a trainable machine learning model which predicts an object boundary box for a time t+k, comparing the predicted object boundary box with a comparison box obtained from the time series of object boundary boxes for the time t+k, evaluating a deviation between the predicted object boundary box and the comparison box using a predetermined cost function, and optimizing parameters which characterize the behavior of the model.
1 . A method for training a machine learning model to predict motion of a detected object, the method comprising:
providing a time series of frames of first sensor data recorded by physical observation of at least one object;
providing a time series of first object boundary boxes encompassing the at least one object in the time series of frames of the first sensor data; and
training the machine learning model to predict motion of the at least one object using the time series of frames of the first sensor data and the time series of the first object boundary boxes as training data, the training comprising:
determining, using the machine learning model, a predicted object boundary box for a time “t+k” based on (i) the object boundary box at a time “t” and (ii) a first history of sensor data from the time series of frames of the first sensor data prior to the time “t” and/or a first history of object boundary boxes from the time series of the first object boundary boxes prior to the time “t”, where is the time “t+k” is after the time “t”;
comparing the predicted object boundary box with a comparison object boundary box obtained from the time series of the first object boundary boxes at the time “t+k”;
evaluating a deviation between the predicted object boundary box and the comparison object boundary box using a predetermined cost function; and
optimizing parameters which characterize a behavior of the machine learning model based on the deviation.
2 . The method according to claim 1 , wherein at least a portion of the first object boundary boxes in the time series of the first object boundary boxes is determined by an object detector from a respective time-corresponding frame from the time series of frames of the first sensor data.
3 . The method according to claim 1 , wherein the machine learning model determines the predicted object boundary box using at least one physical movement model.
4 . The method according to claim 3 , wherein the physical movement model relies on at least one of an assumption of a constant velocity, a constant velocity and a direction of travel, and a constant velocity and rate of rotation.
5 . The method according to claim 1 , wherein the time series of frames of the first sensor data includes radar, LiDAR, camera, and/or ultrasonic sensor data.
6 . The method according to claim 1 , wherein:
at least one object boundary box in the time series of the first object boundary boxes is annotated with at least one object parameter indicative of a pose, linear velocity, radial velocity, radial acceleration, two-dimensional velocity, two-dimensional acceleration, angular velocity, and/or rotational rate of the object, and/or said object parameter is regressed by an object detector; and
the machine learning model receives the at least one object parameter as an additional input.
7 . The method according to claim 6 , wherein:
a predicted object parameter for the time “t+k” is determined using the machine learning model and/or is determined using the predicted object boundary box determined by the machine learning model;
evaluating a further deviation between the predicted object parameter for the time “t+k” and the object parameter regressed by the object detector at the time “t+k” using a further cost function; and
optimizing the parameters that characterize the behavior of the object detector based on the further deviation.
8 . The method according to claim 1 , wherein:
additional movement parameters are estimated using sensor data including a plurality of reflections of sensor signals off of the respective object, such that velocity information beyond a radial velocity of the object is obtained; and
the machine learning model receives the additional movement parameters as input to the machine learning model.
9 . The method according to claim 1 , wherein additional information regarding ego motion of sensors that measured the time series of frames of the first sensor data are used to perform ego motion compensation, and/or are supplied to the machine learning model as inputs.
10 . The method according to claim 1 , further comprising, after the training the machine learning model:
supplying an object boundary box to the machine learning model at a current time along with at least one of a second history of sensor data from a time series of frames of second sensor data prior to the current time and a second history of object boundary boxes from a time series of second object boundary boxes prior to the current time; and
determining, using object boundary boxes predicted using the machine learning model, a predicted future movement of at least one object.
11 . The method according to claim 10 , wherein the second history of sensor data prior to the current time are recorded by at least one sensor accommodated on or included in a vehicle.
12 . The method according to claim 11 , further comprising:
generating a control signal based on the predicted future movement of the at least one object; and
controlling the vehicle using the control signal so as to cause the vehicle to brake, accelerate, continue at constant velocity, and/or take an evasive maneuver.
13 . The method according to claim 1 , wherein the method is implemented by a computer that executes a computer program.
14 . A non-transitory machine-readable data storage medium that stores a computer program having instructions that, when executed on a computer, cause the computer to perform the method according to claim 1 .
15 . A computer comprising:
at least one memory; and
at least one processor operably connected to the at least one memory, the at least one processor being configured to execute instructions of a computer program to perform the method according to claim 1 .