Object tracking method and recording medium
View Patent ↗An object tracking method includes: predicting a prediction distribution which is a distribution of each of prediction pose locations of first objects in a current frame by inputting pose time-sequence data to a pose predictor; estimating a set of locations of each of second objects in the current frame by inputting the current frame to a pose estimator; obtaining identification information indicating a one-to-one correspondence between the second objects and the first objects and location information of the second objects by matching the set of locations and the prediction distribution, to output the obtained identification information and location information; and updating the pose time-sequence data by adding, to the pose time-sequence data and based on the obtained identification information and location information, data which includes coordinates indicating a pose of each of the first objects and is obtained from the predicted prediction distribution.
1. An object tracking method performed by a computer for tracking an object in a video, the object tracking method comprising:
predicting, by a pose predictor, a prediction distribution by inputting pose time-sequence data to the pose predictor, the pose predictor including a first neural network and performing pose prediction which takes uncertainty into consideration, the prediction distribution being a distribution of each of prediction pose locations of one or more first objects in a current frame, the pose time-sequence data including coordinates indicating a pose of each of the one or more first objects in at least two previous frames in time sequence;
estimating, by a second neural network, a set of locations of each of one or more second objects in the current frame by inputting the current frame to the second neural network;
obtaining identification information and location information by matching the set of locations and the prediction distribution, to output the identification information and location information obtained, the identification information indicating a one-to-one correspondence between the one or more second objects and the one or more first objects, the location information indicating a location of a pose of each of the one or more second objects; and
updating the pose time-sequence data by adding, to the pose time-sequence data, data including coordinates indicating a pose of each of the one or more first objects based on the identification information and the location information obtained in the obtaining, the data being obtained from the prediction distribution predicted in the predicting.
2. The object tracking method according to claim 1 , wherein
in the obtaining,
the matching is performed to calculate a pair of the prediction distribution and the set of the locations, the pair attaining a matching score that is at least a predetermined value, and the identification information and the location information are obtained based on the pair calculated and are output.
3. The object tracking method according to claim 2 , wherein
in the obtaining,
the matching performed to calculate the pair is bipartite matching.
4. The object tracking method according to claim 1 , wherein
in the predicting,
the first neural network predicts the prediction distribution using one or more models for taking the uncertainty into consideration, by causing each of the one or more models to predict the prediction pose locations from the pose time-sequence data.
5. The object tracking method according to claim 4 , wherein
the prediction distribution takes the uncertainty into consideration, includes the prediction pose locations predicted, and is expressed by dispersion of a Gaussian distribution.
6. The object tracking method according to claim 4 , wherein
the first neural network includes a recurrent neural network and a predetermined neural network that creates the one or more models.
7. The object tracking method according to claim 6 , wherein
the one or more models are created by applying dropout to the predetermined neural network and inactivating one or more nodes constituting the predetermined neural network.
8. A non-transitory computer-readable recording medium for use in a computer, the recording medium having a computer program recorded thereon for causing the computer to execute:
predicting, by a pose predictor, a prediction distribution by inputting pose time-sequence data to the pose predictor, the pose predictor including a first neural network and performing pose prediction which takes uncertainty into consideration, the prediction distribution being a distribution of each of prediction pose locations of one or more first objects in a current frame, the pose time-sequence data including coordinates indicating a pose of each of the one or more first objects in at least two previous frames in time sequence;
estimating, by a second neural network, a set of locations of each of one or more second objects in the current frame by inputting the current frame to the second neural network;
obtaining identification information and location information by matching the set of locations and the prediction distribution, to output the identification information and location information obtained, the identification information indicating a one-to-one correspondence between the one or more second objects and the one or more first objects, the location information indicating a location of a pose of each of the one or more second objects; and
updating the pose time-sequence data by adding, to the pose time-sequence data, data including coordinates indicating a pose of each of the one or more first objects based on the identification information and the location information obtained in the obtaining, the data being obtained from the prediction distribution predicted in the predicting.