Multiple targets—tracking method and apparatus, device and storage medium

1. A multiple targets-tracking method, comprising:

obtaining a to-be-processed current image, inputting the current image into a convolutional neural network model obtained by pre-training, to perform detection of multiple targets;

respectively extracting a feature vector of each detected target from features of the current image generated by a pre-selected convolutional layer of the convolutional neural network model during the detection of multiple targets; and

respectively calculating a similarity between the feature vector of each target in current image and a feature vector of each target in previous images, completing association of the same target between different image frames according to calculation results, and allocating a same tracking serial number to the same target between different image frames.

2. The method according to claim 1 , wherein

the obtaining a to-be-processed current image, inputting the current image into a convolutional neural network model obtained by pre-training, to perform detection of multiple targets further comprises:

obtaining images as training data, and respectively obtaining a notation result of each target in each frame of images as the training data;

obtaining the convolutional neural network model by training according to the images as the training data and the notation results of each target therein.

3. The method according to claim 1 , wherein

the respectively extracting the feature vector of each detected target from features of the current image generated by a pre-selected convolutional layer comprises:

respectively performing the following processing for each detected target:

finding a corresponding area of the target in the selected convolutional layer;

extracting convolutional features corresponding to the area in a roi pooling manner;

normalizing the extracted convolutional features to obtain the feature vector of the target.

4. The method according to claim 1 , wherein

the respectively calculating a similarity between the feature vector of each target in current image and the feature vector of each target in previous images comprises:

with respect to each detected target, respectively calculating the similarity between the feature vector of the target and the feature vector of each target in previous images according to a similarity estimating model obtained by pre-training.

5. The method according to claim 4 , wherein

before obtaining a to-be-processed current image, inputting the current image into a convolutional neural network model obtained by pre-training, to perform detection of multiple targets, the method further comprises:

obtaining each target as the training data and corresponding feature vectors; and

obtaining the similarity estimating model by training according to a principle that a similarity between feature vectors from the same target is larger than a similarity between feature vectors from different targets.

6. The method according to claim 5 , wherein

the similarity estimating model is:

similarity( x,z )= x T θz,

wherein x and z respectively represent two feature vectors;

x T is the transpose of x;

θ represents a similarity parameter obtained from the training;

similarity(x,z) represents the similarity resulting from the calculation.

7. A computer device, comprising a memory, a processor and a computer program which is stored on the memory and runs on the processor, wherein the processor, upon executing the program, implements the following operation:

obtaining a to-be-processed current image, inputting the current image into a convolutional neural network model obtained by pre-training, to perform detection of multiple targets;

respectively extracting a feature vector of each detected target from features of the current image generated by a pre-selected convolutional layer of the convolutional neural network model during the detection of multiple targets; and

respectively calculating a similarity between the feature vectors of each target in current image and a feature vector of each target in previous images, completing association of the same target between different image frames according to calculation results, and allocating a same tracking serial number to the same target between different image frames.

8. The computer device according to claim 7 , wherein

the obtaining a to-be-processed current image, inputting the current image into a convolutional neural network model obtained by pre-training, to perform detection of multiple targets further comprises:

obtaining images as training data, and respectively obtaining a notation result of each target in each frame of images as the training data; and

obtaining the convolutional neural network model by training according to the images as the training data and the notation results of each target therein.

9. The computer device according to claim 7 , wherein

the respectively extracting the feature vectors of each detected target from features of the current image generated by a pre-selected convolutional layer comprises:

respectively performing the following processing for each detected target:

finding a corresponding area of the target in the selected convolutional layer;

extracting convolutional features corresponding to the area in a roi pooling manner; and

normalizing the extracted convolutional features to obtain the feature vector of the target.

10. The computer device according to claim 7 , wherein

the respectively calculating a similarity between the feature vector of each target in current image and the feature vectors of each target in previous images comprises:

with respect to each detected target, respectively calculating the similarity between the feature vector of the target and the feature vector of each target in previous images according to a similarity estimating model obtained by pre-training.

11. The computer device according to claim 10 , wherein

before obtaining a to-be-processed current image, inputting the current image into a convolutional neural network model obtained by pre-training, to perform detection of multiple targets, the operation further comprises:

obtaining each target as the training data and corresponding feature vectors; and

obtaining the similarity estimating model by training according to a principle that a similarity between feature vectors from the same target is larger than a similarity between feature vectors from different targets.

12. The computer device according to claim 11 , wherein

the similarity estimating model is:

similarity( x,z )= x T θz,

wherein x and z respectively represent two feature vectors;

x T is the transpose of x;

θ represents a similarity parameter obtained from the training;

similarity(x,z) represents the similarity resulting from the calculation.

13. A non-transitory computer-readable storage medium on which a computer program is stored, wherein the program, when executed by a processor, implements the following operation:

obtaining a to-be-processed current image, inputting the current image into a convolutional neural network model obtained by pre-training, to perform detection of multiple targets;

respectively extracting a feature vectors of each detected target from features of the current image generated by a pre-selected convolutional layer of the convolutional neural network model during the detection of multiple targets; and

respectively calculating a similarity between the feature vectors of each target in current image and a feature vectors of each target in previous images, completing association of the same target between different image frames according to calculation results, and allocating a same tracking serial number to the same target between different image frames.

14. The non-transitory computer-readable storage medium according to claim 13 , wherein

the obtaining a to-be-processed current image, inputting the current image into a convolutional neural network model obtained by pre-training, to perform detection of multiple targets further comprises:

obtaining images as training data, and respectively obtaining a notation result of each target in each frame of images as the training data; and

obtaining the convolutional neural network model by training according to the images as the training data and the notation results of each target therein.

15. The non-transitory computer-readable storage medium according to claim 13 , wherein

the respectively extracting the feature vector of each detected target from features of the current image generated by a pre-selected convolutional layer comprises:

respectively performing the following processing for each detected target:

finding a corresponding area of the target in the selected convolutional layer;

extracting convolutional features corresponding to the area in a roi pooling manner; and

normalizing the extracted convolutional features to obtain the feature vector of the target.

16. The non-transitory computer-readable storage medium according to claim 13 , wherein

the respectively calculating a similarity between the feature vector of each target in current image and the feature vector of each target in previous images comprises:

with respect to each detected target, respectively calculating the similarity between the feature vectors of the target and the feature vectors of each target in previous images according to a similarity estimating model obtained by pre-training.

17. The non-transitory computer-readable storage medium according to claim 16 , wherein

before obtaining a to-be-processed current image, inputting the current image into a convolutional neural network model obtained by pre-training, to perform detection of multiple targets, the operation further comprises:

obtaining each target as the training data and corresponding feature vectors; and

obtaining the similarity estimating model by training according to a principle that a similarity between feature vectors from the same target is larger than a similarity between feature vectors from different targets.

18. The non-transitory computer-readable storage medium according to claim 17 , wherein

the similarity estimating model is:

similarity( x,z )= x T θz,

wherein x and z respectively represent two feature vectors;

x T is the transpose of x;

θ represents a similarity parameter obtained from the training;

similarity(x,z) represents the similarity resulting from the calculation.