Method for common detecting, tracking and classifying of objects
A method for machine-based training of a computer-implemented network for common detecting, tracking, and classifying of at least one object in a video image sequence having a plurality of successive individual images. A combined error may be determined during the training, which error results from the errors of the determining of the class identification vector, determining of the at least one identification vector, the determining of the specific bounding box regression, and the determining of the inter-frame regression.
1. A method for common detecting, tracking, and classifying of at least one object in a video image sequence having a multiplicity of successive frames by means of a trained computer-implemented network; wherein the method comprises:
receiving a first frame and a succeeding second frame;
detecting at least one object in the first frame and the succeeding second frame;
selecting an object from the first and second frames;
ascertaining at least one classification vector and a position for the object from the first and second frames;
ascertaining an association value on the basis of the ascertained classification vector and the position comprising providing a relative weighting between the at least one ascertained classification vector and the position by an already trained network during its execution in detecting, tracking and classifying, said relative weighting being dependent on the time between the received first and succeeding second frames;
generating a temporarily consistent and unique identification vector of the at least one object for each frame in response to the ascertained association value.
2. The method as claimed in claim 1 , wherein the detecting at least one object comprises:
generating a bounding box surrounding the at least one object;
generating a prediction for the bounding box from the first frame to the second frame;
generating a velocity vector for the bounding box of the first frame.
3. The method as claimed in claim 2 , wherein a bounding box is provided for each of the at least one object.
4. The method as claimed in claim 2 , wherein the selecting comprises at least one of:
selecting the bounding box of the first frame and selecting the bounding box of the second frame;
selecting the prediction and selecting the bounding box of the second frame; and
selecting the velocity vector and selecting the bounding box of the second frame.
5. The method as claimed in claim 1 , wherein the ascertaining at least one classification vector comprises:
acquiring features of the object;
calculating a unique feature vector from the acquired features; and
classifying the object from a group of predefined classes on the basis of the acquired features or on the basis of calculated feature vector.
6. The method as claimed in claim 1 , wherein the relative weighting between the ascertained classification vector and the position rises with increasing time or a falling frame rate between the first and second frames.
7. The method as claimed in claim 1 , wherein the generating a temporarily consistent and unique identification vector comprises a Hungarian combinatorial optimization method.
8. The method as claimed in claim 1 , wherein the unique identification of an object of a second frame that is not assignable to any object of a first frame is compared with the identification of an object of a third frame temporally proceeding the first frame.
9. A system for classifying objects on a computer which comprises:
a memory and one or more processors configured to perform the method as claimed in claim 1 .
10. A computer program product stored on a non-transitory computer readable medium and having instructions which, when executed on one or more processors, carry out the method as claimed in claim 1 .