Method for estimating motion of object based on vision sensor and object motion estimating device using the same
There is provided a method for estimating a motion of an object based on a vision sensor. The method includes steps of: (a) inputting a t-th vision sensor data and a t-th vehicle system data into a data preprocessing module, to thereby instruct the data preprocessing module to preprocess a t-th vision sensor data and a t-th vehicle system data and thus to generate a t-th vector for estimating object motion, and (b) inputting the t-th vector for estimating the motion of the object into a deep learning based sequential regression network, to thereby instruct the sequential regression network to perform a recurrent learning operation on the t-th vector for estimating object motion and thus to generate a t-th prediction object motion data which is a result of predicting object motion corresponding to the t-th image frame.
1 . A method for estimating a motion of an object based on a vision sensor, comprising steps of:
(a) in response to sequentially acquiring a vision sensor data and a vehicle system data, inputting, by an object motion estimating device, a t-th vision sensor data corresponding to a t-th image frame, which is a current image frame, and a t-th vehicle system data into a data preprocessing module, to thereby instruct the data preprocessing module to preprocess the t-th vision sensor data and the t-th vehicle system data and thus to generate a t-th vector for estimating the motion of the object, wherein the vision sensor data includes a raw object relative distance and an object heading angle, wherein the raw object relative distance is a distance from a vehicle to the object detected through a deep learning based detection model from at least one image captured by at least one camera installed in the vehicle, wherein the object heading angle is a heading angle of the object, and wherein the vehicle system data is related to a data detected from a system of the vehicle;
(b) inputting, by the object motion estimating device, the t-th vector for estimating the motion of the object into a deep learning based sequential regression network, to thereby instruct the sequential regression network to perform a recurrent learning operation on the t-th vector for estimating the motion of the object and thus to generate a t-th prediction object motion data which is a result of predicting the motion of the object corresponding to the t-th image frame; and
(c) inputting, by the object motion estimating device, the t-th prediction object motion data and the t-th vehicle system data into a sequential filtering network, to thereby instruct the sequential filtering network to (i) generate, through a state prediction model, a t-th prediction compensation object motion data which is a result of predicting a compensation object motion data in the t-th image frame by applying a learning operation to a (t−1)-th compensation object motion data and the t-th vehicle system data, wherein the (t−1)-th compensation object motion data is a result of compensating a (t−1)-th prediction object motion data in a (t−1)-th image frame by referring to the (t−1)-th prediction object motion data and a (t−1)-th vehicle system data, (ii) generate, through a state vector generation module, a t-th prediction difference state vector based on a difference between the t-th prediction object motion data and the (t−1)-th prediction object motion data, a t-th prediction-compensation difference state vector based on a difference between the t-th prediction object motion data and the t-th prediction compensation object motion data, and a t-th compensation difference state vector based on a difference between the t-th prediction compensation object motion data and the (t−1)-th compensation object motion data, (iii) generate a (t_1)-th uncertainty probability value which is a result of estimating an uncertainty of the t-th prediction compensation object motion data by applying a recurrent operation to the t-th compensation difference state vector through a first filtering model, and generate a (t_2)-th uncertainty probability value which is a result of estimating an uncertainty of the t-th prediction object motion data by applying a recurrent operation to the t-th prediction difference state vector and the t-th prediction-compensation difference state vector through a second filtering model, and (iv) generate a t-th compensation object motion data in the t-th image frame by applying a t-th feedback gain to the t-th prediction-compensation difference state vector and thus by compensating the t-th prediction compensation object motion data, wherein the t-th feedback gain is generated by referring to the (t_1)-th uncertainty probability value and the (t_2)-th uncertainty probability value.
2 . The method of claim 1 , wherein, at the step of (c), the object motion estimating device instructs the sequential filter network to (i) further generate a t-th object distance state vector based on a t-th prediction object relative distance included in the t-th prediction object motion data through the state vector generation module, (ii) generate the (t_1)-th uncertainty probability value which is a result of estimating an uncertainty of the t-th prediction compensation object motion data by applying a recurrent operation to the t-th compensation difference state vector and the t-th object distance state vector through the first filtering model, and (iii) generate the (t_2)-th uncertainty probability value which is a result of estimating an uncertainty of the t-th prediction object motion data by applying a recurrent operation to the t-th prediction difference state vector, the t-th prediction-compensation difference state vector, and the t-th object distance state vector through the second filtering model.
3 . The method of claim 1 , wherein, at the step of (c), the t-th vehicle system data includes a t-th vehicle velocity which is a velocity of the vehicle in the t-th image frame, a t-th vehicle acceleration which is an acceleration of the vehicle in the t-th image frame, a t-th vehicle yaw rate which is a yaw rate of the vehicle in the t-th image frame, and a t-th time difference which is a difference between a time at which the (t−1)-th image frame is acquired and a time at which the t-th image frame is acquired.
4 . The method of claim 1 , wherein, at the step of (c), the state prediction model comprises an exercise model including at least one of a constant velocity model and a constant acceleration model.
5 . The method of claim 1 , wherein, at the step of (c), each of the first filtering model and the second filtering model comprises one of an RNN (Recurrent Neural Network), an LSTM (Long Short-Term Memory), and a GRU (Gated Recurrent Unit).
6 . A method for estimating a motion of an object based on a vision sensor, comprising steps of:
(a) in response to sequentially acquiring a vision sensor data and a vehicle system data, inputting, by an object motion estimating device, a t-th vision sensor data corresponding to a t-th image frame, which is a current image frame, and a t-th vehicle system data into a data preprocessing module, to thereby instruct the data preprocessing module to preprocess the t-th vision sensor data and the t-th vehicle system data and thus to generate a t-th vector for estimating the motion of the object, wherein the vision sensor data includes a raw object relative distance and an object heading angle, wherein the raw object relative distance is a distance from a vehicle to the object detected through a deep learning based detection model from at least one image captured by at least one camera installed in the vehicle, wherein the object heading angle is a heading angle of the object, and wherein the vehicle system data is related to a data detected from a system of the vehicle; and
(b) inputting, by the object motion estimating device, the t-th vector for estimating the motion of the object into a deep learning based sequential regression network, to thereby instruct the sequential regression network to perform a recurrent learning operation on the t-th vector for estimating the motion of the object and thus to generate a t-th prediction object motion data which is a result of predicting the motion of the object corresponding to the t-th image frame;
wherein, at the step of (a), the object motion estimating device instructs the data preprocessing module to generate, as the t-th vector for estimating the motion of the object, a t-th vector for estimating a relative distance of the object, a t-th vector for estimating an absolute velocity of the object, and a t-th vector for estimating an absolute acceleration of the object by referring to the t-th vision sensor data and the t-th vehicle system data, wherein the t-th vector for estimating the relative distance of the object includes a t-th data for estimating the relative distance of the object, wherein the t-th vector for estimating the absolute velocity of the object includes a t-th data for estimating the absolute velocity of the object, and wherein the t-th vector for estimating the absolute acceleration of the object includes a t-th data for estimating the absolute acceleration of the object; and
wherein, at the step of (b), the object motion estimating device instructs the sequential regression network to (i) generate a t-th prediction object relative distance which is a result of predicting the relative distance of the object with respect to the vehicle in the t-th image frame by applying a recurrent operation to the t-th vector for estimating the relative distance of the object through an object relative distance estimating model, (ii) generate a t-th prediction object absolute velocity which is a result of predicting the absolute velocity of the object in the t-th image frame by applying a recurrent operation to the t-th vector for estimating the absolute velocity of the object through an object absolute velocity estimating model, (iii) generate a t-th prediction object absolute acceleration which is a result of predicting the absolute acceleration of the object in the t-th image frame by applying a recurrent operation to the t-th vector for estimating the absolute acceleration of the object through an object absolute acceleration estimating model, and thus (iv) generate the t-th prediction object motion data which includes the t-th prediction object relative distance, the t-th prediction object absolute velocity, and the t-th prediction object absolute acceleration.
7 . A method for estimating a motion of an object based on a vision sensor, comprising steps of:
(a) in response to sequentially acquiring a vision sensor data and a vehicle system data, inputting, by an object motion estimating device, a t-th vision sensor data corresponding to a t-th image frame, which is a current image frame, and a t-th vehicle system data into a data preprocessing module, to thereby instruct the data preprocessing module to preprocess the t-th vision sensor data and the t-th vehicle system data and thus to generate a t-th vector for estimating the motion of the object, wherein the vision sensor data includes a raw object relative distance and an object heading angle, wherein the raw object relative distance is a distance from a vehicle to the object detected through a deep learning based detection model from at least one image captured by at least one camera installed in the vehicle, wherein the object heading angle is a heading angle of the object, and wherein the vehicle system data is related to a data detected from a system of the vehicle; and
(b) inputting, by the object motion estimating device, the t-th vector for estimating the motion of the object into a deep learning based sequential regression network, to thereby instruct the sequential regression network to perform a recurrent learning operation on the t-th vector for estimating the motion of the object and thus to generate a t-th prediction object motion data which is a result of predicting the motion of the object corresponding to the t-th image frame;
wherein, at the step of (a), the object motion estimating device instructs the data preprocessing module to generate the t-th vector for estimating the motion of the object which includes at least part of a t-th raw object relative distance, a t-th raw object relative velocity which is a relative velocity of the object with respect to the vehicle in the t-th image frame, a t-th raw object absolute velocity which is an absolute velocity of the object in the t-th image frame, a t-th vehicle velocity which is a velocity of the vehicle in the t-th image frame, a t-th vehicle acceleration which is an acceleration of the vehicle in the t-th image frame, a t-th vehicle yaw rate which is a yaw rate of the vehicle in the t-th image frame, a t-th object heading angle, and a t-th time difference which is a difference between a time at which a (t−1)-th image frame is acquired and a time at which the t-th image frame is acquired; and
wherein the t-th raw object relative distance, the t-th raw object relative velocity, the t-th raw object absolute velocity, the t-th vehicle velocity, the t-th vehicle acceleration, the t-th vehicle yaw rate, the t-th object heading angle, and the t-th time difference are included in the t-th vision sensor data and the t-th vehicle system data, or are calculated by referring to the t-th vision sensor data and the t-th vehicle system data.
8 . The method of claim 7 , wherein, at the step of (a), the object motion estimating device instructs the data preprocessing module to generate, as the t-th vector for estimating the motion of the object, (i) a t-th vector for estimating a relative distance of the object, (ii) a t-th vector for estimating the absolute velocity of the object, and (iii) a t-th vector for estimating an absolute acceleration of the object, by referring to the t-th vision sensor data and the t-th vehicle system data, wherein the t-th vector for estimating the relative distance of the object includes the t-th raw object relative distance, the t-th raw object relative velocity, the t-th object heading angle, and the t-th time difference, wherein the t-th vector for estimating the absolute velocity of the object includes the t-th raw object relative distance, the t-th raw object absolute velocity, the t-th vehicle velocity, the t-th vehicle yaw rate, the t-th object heading angle, and the t-th time difference, and wherein the t-th vector for estimating the absolute acceleration of the object includes the t-th raw object relative distance, the t-th raw object absolute velocity, the t-th vehicle velocity, the t-th vehicle acceleration, the t-th vehicle yaw rate, the t-th object heading angle, and the t-th time difference; and
wherein, at the step of (b), the object motion estimating device instructs the sequential regression network to (i) generate a t-th prediction object relative distance which is a result of predicting the relative distance of the object with respect to the vehicle in the t-th image frame by applying a recurrent operation to the t-th vector for estimating the relative distance of the object through an object relative distance estimating model, (ii) generate a t-th prediction object absolute velocity which is a result of predicting the absolute velocity of the object in the t-th image frame by applying a recurrent operation to the t-th vector for estimating the absolute velocity of the object through an object absolute velocity estimating model, (iii) generate a t-th prediction object absolute acceleration which is a result of predicting the absolute acceleration of the object in the t-th image frame by applying a recurrent operation to the t-th vector for estimating the absolute acceleration of the object through an object absolute acceleration estimating model, and thus (iv) generate the t-th prediction object motion data which includes the t-th prediction object relative distance, the t-th prediction object absolute velocity, and the t-th prediction object absolute acceleration.
9 . The method of claim 8 , wherein, at the step of (a), the object motion estimating device instructs the data preprocessing module to generate the t-th vector for estimating the motion of the object, such that each of the t-th vector for estimating the relative distance of the object, the t-th vector for estimating the absolute velocity of the object, and the t-th vector for estimating the absolute acceleration of the object further includes a t-th vehicle pitch which is a pitch of the vehicle in the t-th image frame, and a t-th vehicle roll which is a roll of the vehicle in the t-th image frame.
10 . The method of claim 1 , wherein, at the step of (b), the sequential regression network comprises one of an RNN (Recurrent Neural Network), an LSTM (Long Short-Term Memory), and a GRU (Gated Recurrent Unit).
11 . An object motion estimating device for estimating a motion of an object based on a vision sensor, comprising:
at least one memory which saves instructions for estimating the motion of the object based on the vision sensor; and
at least one processor configured to execute an operation for estimating the motion of the object based on the vision sensor according to the instructions saved in the memory to perform processes of: (I) in response to sequentially acquiring a vision sensor data and a vehicle system data, inputting a t-th vision sensor data corresponding to a t-th image frame, which is a current image frame, and a t-th vehicle system data into a data preprocessing module, to thereby instruct the data preprocessing module to preprocess the t-th vision sensor data and the t-th vehicle system data and thus to generate a t-th vector for estimating the motion of the object, wherein the vision sensor data includes a raw object relative distance and an object heading angle, wherein the raw object relative distance is a distance from a vehicle to the object detected through a deep learning based detection model from at least one image captured by at least one camera installed in the vehicle, wherein the object heading angle is a heading angle of the object, and wherein the vehicle system data is related to a data detected from a system of the vehicle; (II) inputting the t-th vector for estimating the motion of the object into a deep learning based sequential regression network, to thereby instruct the sequential regression network to perform a recurrent learning operation on the t-th vector for estimating the motion of the object and thus to generate a t-th prediction object motion data which is a result of predicting the motion of the object corresponding to the t-th image frame; and (III) inputting the t-th prediction object motion data and the t-th vehicle system data into a sequential filtering network, to thereby instruct the sequential filtering network to (i) generate, through a state prediction model, a t-th prediction compensation object motion data which is a result of predicting a compensation object motion data in the t-th image frame by applying a learning operation to a (t−1)-th compensation object motion data and the t-th vehicle system data, wherein the (t−1)-th compensation object motion data is a result of compensating a (t−1)-th prediction object motion data in a (t−1)-th image frame by referring to the (t−1)-th prediction object motion data and a (t−1)-th vehicle system data, (ii) generate, through a state vector generation module, a t-th prediction difference state vector based on a difference between the t-th prediction object motion data and the (t−1)-th prediction object motion data, a t-th prediction-compensation difference state vector based on a difference between the t-th prediction object motion data and the t-th prediction compensation object motion data, and a t-th compensation difference state vector based on a difference between the t-th prediction compensation object motion data and the (t−1)-th compensation object motion data, (iii) generate a (t_1)-th uncertainty probability value which is a result of estimating an uncertainty of the t-th prediction compensation object motion data by applying a recurrent operation to the t-th compensation difference state vector through a first filtering model, and generate a (t_2)-th uncertainty probability value which is a result of estimating an uncertainty of the t-th prediction object motion data by applying a recurrent operation to the t-th prediction difference state vector and the t-th prediction-compensation difference state vector through a second filtering model, and (iv) generate a t-th compensation object motion data in the t-th image frame by applying a t-th feedback gain to the t-th prediction-compensation difference state vector and thus by compensating the t-th prediction compensation object motion data, wherein the t-th feedback gain is generated by referring to the (t_1)-th uncertainty probability value and the (t_2)-th uncertainty probability value.
12 . The object motion estimating device of claim 11 , wherein, at the process of (III), the processor instructs the sequential filter network to (i) further generate a t-th object distance state vector based on a t-th prediction object relative distance included in the t-th prediction object motion data through the state vector generation module, (ii) generate the (t_1)-th uncertainty probability value which is a result of estimating an uncertainty of the t-th prediction compensation object motion data by applying a recurrent operation to the t-th compensation difference state vector and the t-th object distance state vector through the first filtering model, and (iii) generate the (t_2)-th uncertainty probability value which is a result of estimating an uncertainty of the t-th prediction object motion data by applying a recurrent operation to the t-th prediction difference state vector, the t-th prediction-compensation difference state vector, and the t-th object distance state vector through the second filtering model.
13 . The object motion estimating device of claim 11 , wherein, at the process of (III), the t-th vehicle system data includes a t-th vehicle velocity which is a velocity of the vehicle in the t-th image frame, a t-th vehicle acceleration which is an acceleration of the vehicle in the t-th image frame, a t-th vehicle yaw rate which is a yaw rate of the vehicle in the t-th image frame, and a t-th time difference which is a difference between a time at which the (t−1)-th image frame is acquired and a time at which the t-th image frame is acquired.
14 . The object motion estimating device of claim 11 , wherein, at the process of (III), the state prediction model comprises an exercise model including at least one of a constant velocity model and a constant acceleration model.
15 . The object motion estimating device of claim 11 , wherein, at the process of (III), each of the first filtering model and the second filtering model comprises one of an RNN (Recurrent Neural Network), an LSTM (Long Short-Term Memory), and a GRU (Gated Recurrent Unit).
16 . An object motion estimating device for estimating a motion of an object based on a vision sensor, comprising:
at least one memory which saves instructions for estimating the motion of the object based on the vision sensor; and
at least one processor configured to execute an operation for estimating the motion of the object based on the vision sensor according to the instructions saved in the memory to perform processes of: (I) in response to sequentially acquiring a vision sensor data and a vehicle system data, inputting a t-th vision sensor data corresponding to a t-th image frame, which is a current image frame, and a t-th vehicle system data into a data preprocessing module, to thereby instruct the data preprocessing module to preprocess the t-th vision sensor data and the t-th vehicle system data and thus to generate a t-th vector for estimating the motion of the object, wherein the vision sensor data includes a raw object relative distance and an object heading angle, wherein the raw object relative distance is a distance from a vehicle to the object detected through a deep learning based detection model from at least one image captured by at least one camera installed in the vehicle, wherein the object heading angle is a heading angle of the object, and wherein the vehicle system data is related to a data detected from a system of the vehicle; and (II) inputting the t-th vector for estimating the motion of the object into a deep learning based sequential regression network, to thereby instruct the sequential regression network to perform a recurrent learning operation on the t-th vector for estimating the motion of the object and thus to generate a t-th prediction object motion data which is a result of predicting the motion of the object corresponding to the t-th image frame;
wherein, at the process of (I), the processor instructs the data preprocessing module to generate, as the t-th vector for estimating the motion of the object, a t-th vector for estimating a relative distance of the object, a t-th vector for estimating an absolute velocity of the object, and a t-th vector for estimating an absolute acceleration of the object by referring to the t-th vision sensor data and the t-th vehicle system data, wherein the t-th vector for estimating the relative distance of the object includes a t-th data for estimating the relative distance of the object, wherein the t-th vector for estimating the absolute velocity of the object includes a t-th data for estimating the absolute velocity of the object, and wherein the t-th vector for estimating the absolute acceleration of the object includes a t-th data for estimating the absolute acceleration of the object; and
wherein, at the process of (II), the processor instructs the sequential regression network to (i) generate a t-th prediction object relative distance which is a result of predicting the relative distance of the object with respect to the vehicle in the t-th image frame by applying a recurrent operation to the t-th vector for estimating the relative distance of the object through an object relative distance estimating model, (ii) generate a t-th prediction object absolute velocity which is a result of predicting the absolute velocity of the object in the t-th image frame by applying a recurrent operation to the t-th vector for estimating the absolute velocity of the object through an object absolute velocity estimating model, (iii) generate a t-th prediction object absolute acceleration which is a result of predicting the absolute acceleration of the object in the t-th image frame by applying a recurrent operation to the t-th vector for estimating the absolute acceleration of the object through an object absolute acceleration estimating model, and thus (iv) generate the t-th prediction object motion data which includes the t-th prediction object relative distance, the t-th prediction object absolute velocity, and the t-th prediction object absolute acceleration.
17 . An object motion estimating device for estimating a motion of an object based on a vision sensor, comprising:
at least one memory which saves instructions for estimating the motion of the object based on the vision sensor; and
at least one processor configured to execute an operation for estimating the motion of the object based on the vision sensor according to the instructions saved in the memory to perform processes of: (I) in response to sequentially acquiring a vision sensor data and a vehicle system data, inputting a t-th vision sensor data corresponding to a t-th image frame, which is a current image frame, and a t-th vehicle system data into a data preprocessing module, to thereby instruct the data preprocessing module to preprocess the t-th vision sensor data and the t-th vehicle system data and thus to generate a t-th vector for estimating the motion of the object, wherein the vision sensor data includes a raw object relative distance and an object heading angle, wherein the raw object relative distance is a distance from a vehicle to the object detected through a deep learning based detection model from at least one image captured by at least one camera installed in the vehicle, wherein the object heading angle is a heading angle of the object, and wherein the vehicle system data is related to a data detected from a system of the vehicle; and (II) inputting the t-th vector for estimating the motion of the object into a deep learning based sequential regression network, to thereby instruct the sequential regression network to perform a recurrent learning operation on the t-th vector for estimating the motion of the object and thus to generate a t-th prediction object motion data which is a result of predicting the motion of the object corresponding to the t-th image frame; and
wherein, at the process of (I), the processor instructs the data preprocessing module to generate the t-th vector for estimating the motion of the object which includes at least part of a t-th raw object relative distance, a t-th raw object relative velocity which is a relative velocity of the object with respect to the vehicle in the t-th image frame, a t-th raw object absolute velocity which is an absolute velocity of the object in the t-th image frame, a t-th vehicle velocity which is a velocity of the vehicle in the t-th image frame, a t-th vehicle acceleration which is an acceleration of the vehicle in the t-th image frame, a t-th vehicle yaw rate which is a yaw rate of the vehicle in the t-th image frame, a t-th object heading angle, and a t-th time difference which is a difference between a time at which a (t−1)-th image frame is acquired and a time at which the t-th image frame is acquired, and wherein the t-th raw object relative distance, the t-th raw object relative velocity, the t-th raw object absolute velocity, the t-th vehicle velocity, the t-th vehicle acceleration, the t-th vehicle yaw rate, the t-th object heading angle, and the t-th time difference are included in the t-th vision sensor data and the t-th vehicle system data, or are calculated by referring to the t-th vision sensor data and the t-th vehicle system data.
18 . The object motion estimating device of claim 17 , wherein, at the process of (I), the processor instructs the data preprocessing module to generate, as the t-th vector for estimating the motion of the object, (i) a t-th vector for estimating a relative distance of the object, (ii) a t-th vector for estimating the absolute velocity of the object, and (iii) a t-th vector for estimating an absolute acceleration of the object, by referring to the t-th vision sensor data and the t-th vehicle system data, wherein the t-th vector for estimating the relative distance of the object includes the t-th raw object relative distance, the t-th raw object relative velocity, the t-th object heading angle, and the t-th time difference, wherein the t-th vector for estimating the absolute velocity of the object includes the t-th raw object relative distance, the t-th raw object absolute velocity, the t-th vehicle velocity, the t-th vehicle yaw rate, the t-th object heading angle, and the t-th time difference, and wherein the t-th vector for estimating the absolute acceleration of the object includes the t-th raw object relative distance, the t-th raw object absolute velocity, the t-th vehicle velocity, the t-th vehicle acceleration, the t-th vehicle yaw rate, the t-th object heading angle, and the t-th time difference; and
wherein, at the process of (II), the processor instructs the sequential regression network to (i) generate a t-th prediction object relative distance which is a result of predicting the relative distance of the object with respect to the vehicle in the t-th image frame by applying a recurrent operation to the t-th vector for estimating the relative distance of the object through an object relative distance estimating model, (ii) generate a t-th prediction object absolute velocity which is a result of predicting the absolute velocity of the object in the t-th image frame by applying a recurrent operation to the t-th vector for estimating the absolute velocity of the object through an object absolute velocity estimating model, (iii) generate a t-th prediction object absolute acceleration which is a result of predicting the absolute acceleration of the object in the t-th image frame by applying a recurrent operation to the t-th vector for estimating the absolute acceleration of the object through an object absolute acceleration estimating model, and thus (iv) generate the t-th prediction object motion data which includes the t-th prediction object relative distance, the t-th prediction object absolute velocity, and the t-th prediction object absolute acceleration.
19 . The object motion estimating device of claim 18 , wherein, at the process of (I), the processor instructs the data preprocessing module to generate the t-th vector for estimating the motion of the object, such that each of the t-th vector for estimating the relative distance of the object, the t-th vector for estimating the absolute velocity of the object, and the t-th vector for estimating the absolute acceleration of the object further includes a t-th vehicle pitch which is a pitch of the vehicle in the t-th image frame, and a t-th vehicle roll which is a roll of the vehicle in the t-th image frame.
20 . The object motion estimating device of claim 11 , wherein, at the process of (II), the sequential regression network comprises one of an RNN (Recurrent Neural Network), an LSTM (Long Short-Term Memory), and a GRU (Gated Recurrent Unit).