Device and method for predicting collision area of high speed small object

1 . A device for predicting a collision area of a high speed small object, comprising:

an event camera; and

a processor, coupled to the event camera, wherein

the processor obtains a sparse event stream data through the event camera;

the processor obtains an event stream data corresponding to a cumulative time interval by using the sparse event stream data;

the processor predicts a collision area category by using a temporal spatial feature expression associated with the event stream data, wherein the collision area category corresponds to the high speed small object; and

the processor outputs the collision area category,

wherein the device further comprises a storage medium coupled to the processor, wherein

the processor encodes the event stream data into the temporal spatial feature expression;

wherein the storage medium further stores a neural network, wherein the event stream data comprises a plurality of events, and each of the plurality of events corresponds to an image coordinate, a polarity, and a time stamp, wherein

the processor performs a segmentation operation on the event stream data by using a time blocking size to obtain a plurality of time bins;

the processor obtains a new feature value corresponding to each of the plurality of events by using the time stamp, the time blocking size, and the neural network;

the processor generates the temporal spatial feature expression corresponding to each of the plurality of time bins by using the image coordinate, the polarity, and the new feature value.

2 . The device according to claim 1 , wherein the temporal spatial feature expression is a voxel, wherein the image coordinate corresponds to an event image, and a shape of the voxel is associated with the polarity, the time blocking size, a height of the event image, and a width of the event image.

3 . The device according to claim 1 , wherein the storage medium stores a data optimization model, and the processor accesses and executes the data optimization model, wherein

the data optimization model predicts the collision area category by decoding the temporal spatial feature expression.

4 . The device according to claim 3 , wherein

the processor performs a training operation to train the data optimization model, wherein the training operation is associated with at least one of a curvature algorithm and a pixel number threshold.

5 . The device according to claim 1 , wherein a speed of the high speed small object is greater than 30 m/s, and a size of the high speed small object is less than 2 cm.

6 . A method for predicting a collision area of a high speed small object, adaptable for a device comprising an event camera and a processor, wherein the method for predicting the collision area of the high speed small object comprises:

obtaining, by the processor, a sparse event stream data through the event camera;

obtaining, by the processor, an event stream data corresponding to a cumulative time interval by using the sparse event stream data;

predicting, by the processor, a collision area category by using a temporal spatial feature expression associated with the event stream data, wherein the collision area category corresponds to the high speed small object; and

outputting, by the processor, the collision area category,

wherein the device further comprises a storage medium, wherein predicting the collision area category by using the temporal spatial feature expression associated with the event stream data comprises:

encoding, by the processor, the event stream data into the temporal spatial feature expression;

wherein the storage medium further stores a neural network, the event stream data comprises a plurality of events, and each of the plurality of events corresponds to an image coordinate, a polarity, and a time stamp, wherein encoding the event stream data into the temporal spatial feature expression comprises:

performing, by the processor, a segmentation operation on the event stream data by using a time blocking size to obtain a plurality of time bins;

obtaining, by the processor, a new feature value corresponding to each of the plurality of events by using the time stamp, the time blocking size, and the neural network; and

generating, by the processor, the temporal spatial feature expression corresponding to each of the plurality of time bins by using the image coordinate, the polarity, and the new feature value.

7 . The method according to claim 6 , wherein the temporal spatial feature expression is a voxel, the image coordinate corresponds to an event image, and a shape of the voxel is associated with the polarity, the time blocking size, a height of the event image, and a width of the event image.

8 . The method according to claim 6 , wherein the storage medium stores a data optimization model, and the processor accesses and executes the data optimization model, wherein predicting the collision area category by using the temporal spatial feature expression associated with the event stream data comprises:

predicting, by the data optimization model, the collision area category by decoding the temporal spatial feature expression.

9 . The method according to claim 8 , further comprising:

performing, by the processor, a training operation to train the data optimization model, wherein the training operation is associated with at least one of a curvature algorithm and a pixel number threshold.

10 . The method according to claim 6 , wherein a speed of the high speed small object is greater than 30 m/s, and a size of the high speed small object is less than 2 cm.