Image capture direction recognition method and server, surveillance method and system and image capture device

1. A method comprising:

extracting deep features of a target image captured by a camera;

determining, based on degrees of matching between the deep features of the target image and deep features of a plurality of reference images, a matched reference image of the target image;

obtaining a coordinate position relationship between the matched reference image and the target image; and

calculating, using the coordinate position relationship and direction information of the matched reference image, an image capture direction of the camera at the time of capturing the target image;

wherein determining, based on degree of matching between the deep features of the target image and deep features of a plurality of reference images, a matched reference image of the target image comprises:

calculating respective degrees of matching between deep features of the target image and of the plurality of reference images; and

designating the reference image corresponding to the deep features with the greatest degree of matching as the reference image matching the target image.

2. The method of claim 1 , wherein the plurality of reference images comprises:

a predetermined plurality of reference images having at least one of direction features and time features.

3. The method of claim 2 , wherein the plurality of reference images having at least one of direction and time features is determined by steps comprising:

selecting a plurality of feature images having at least one of direction features and time features from historical sample images;

performing cluster analysis on deep features of the historical sample images with respective deep features of the plurality of feature images designated as cluster centers, assigning the historical sample images to a plurality of clusters corresponding to the cluster centers;

determining an image from each cluster as a reference image having corresponding at least one of direction features and time features.

4. The method of claim 3 , wherein determining an image from each cluster as a reference image having corresponding at least one of direction features and time features comprises:

performing a respective mean calculation on deep features of images of each cluster, designating a calculated mean as a respective new cluster center; and

performing iterative cluster analysis on deep features of the historical sample images based on a plurality of new cluster centers until reaching a preset convergence condition, designating the images corresponding to the cluster centers upon reaching the preset convergence condition as reference images;

wherein the preset convergence condition comprises a degree of difference between deep features of historical sample images of each cluster and the respective corresponding cluster center of the cluster being less than or equal to a preset degree of difference value.

5. The method of claim 3 , wherein determining an image from each cluster as a reference image having corresponding at least one of direction features and time features comprises:

calculating respective degrees of difference between deep features of images of each cluster;

selecting from each cluster the image having a smallest sum of degrees of difference from deep features of other images of the same cluster as a reference image.

6. The method of claim 1 , wherein the degree of matching comprises one of:

Euclidean distance;

cosine distance; and

Manhattan distance.

7. The method of claim 1 , wherein obtaining a coordinate position relationship between the matched reference image and the target image comprises:

extracting point features of the matched reference image and the target image;

performing point matching between feature points of the matched reference image and feature points of the target image;

calculating a coordinate position relationship between the matched point pairs, obtaining a homography transformation matrix between the matched reference image and the target image, and designating a conversion relationship of the homography transformation matrix as the coordinate position relationship.

8. The method of claim 1 , wherein direction information comprises:

pan, tilt, and zoom data of the camera at the time of capturing an image.

9. The method of claim 1 , wherein calculating an image capture direction of the camera at the time of capturing the target image comprises:

performing displacement processing on direction information of the matched reference image, based on displacement parameters determined from the coordinate position relationship, to obtain the image capture direction.

10. A server, comprising a processor and a memory, the memory storing computer program instructions executable by the processor to:

extract deep features of a target image captured by a camera;

determine a matched reference image of the target image based on degrees of matching between deep features of the target image and deep features of a plurality of reference images;

obtain a coordinate position relationship between the matched reference image and the target image; and

calculate the image capture direction of the camera at the time of capturing the target image using the coordinate position relationship and direction information of the matched reference image;

wherein determining, based on degree of matching between the deep features of the target image and deep features of a plurality of reference images, a matched reference image of the target image comprises:

calculating respective degrees of matching between deep features of the target image and of the plurality of reference images;

designating the reference image corresponding to the deep features with the greatest degree of matching as the reference image matching the target image.

11. The server of claim 10 , wherein the plurality of reference images comprises:

a predetermined plurality of reference images having at least one of direction features and time features.

12. The server of claim 11 , wherein the plurality of reference images having at least one of direction features and time features is determined by steps comprising:

selecting a plurality of feature images having at least one of direction features and time features from historical sample images;

performing cluster analysis on deep features of the historical sample images with respective deep features of the plurality of feature images designated as cluster centers, assigning the historical sample images to a plurality of clusters corresponding to the cluster centers;

determining an image from each cluster as a reference image having corresponding at least one of direction features and time features.

13. The server of claim 12 , wherein determining an image from each cluster as a reference image having corresponding at least one of direction features and time features comprises:

performing a respective mean calculation on deep features of images of each cluster, designating a calculated mean as a respective new cluster center; and

performing iterative cluster analysis on deep features of the historical sample images based on a plurality of new cluster centers until reaching a preset convergence condition, designating the images corresponding to the cluster centers upon reaching the preset convergence condition as reference images;

wherein the preset convergence condition comprises a degree of difference between deep features of historical sample images of each cluster and the respective corresponding cluster center of the cluster being less than or equal to a preset degree of difference value.

14. The server of claim 12 , wherein determining an image from each cluster as a reference image having corresponding at least one of direction features and time features comprises:

calculating respective degrees of difference between deep features of images of each cluster;

selecting from each cluster the image having a smallest sum of degrees of difference from deep features of other images of the same cluster as a reference image.

15. The server of claim 10 , wherein obtaining a coordinate position relationship between the matched reference image and the target image comprises:

extracting point features of the matched reference image and the target image;

performing point matching between feature points of the matched reference image and feature points of the target image;

calculating a coordinate position relationship between the matched point pairs, obtaining a homography transformation matrix between the matched reference image and the target image, and designating a conversion relationship of the homography transformation matrix as the coordinate position relationship.

16. The server of claim 10 , wherein direction information comprises:

pan, tilt, and zoom data of the camera at the time of capturing an image.

17. The server of claim 10 , wherein calculating an image capture direction of the camera at the time of capturing the target image comprises:

performing displacement processing on direction information of the matched reference image, based on displacement parameters determined from the coordinate position relationship, to obtain the image capture direction.

18. A device comprising:

a deep feature extraction module operable to extract deep features of a target image captured by a camera;

a reference image determination module operable to determine a matched reference image of the target image, based on degrees of matching between deep features of the target image and deep features of a plurality of reference images;

a coordinate position relationship acquisition module operable to obtain a coordinate position relationship between the matched reference image and the target image; and

an image capture direction computation module operable to calculate an image capture direction of the camera at the time of capturing the target image, using the coordinate position relationship and direction information of the matched reference image;

wherein the reference image determination module is operable to determine, based on degree of matching between the deep features of the target image and deep features of a plurality of reference images, a matched reference image of the target image by:

calculating respective degrees of matching between deep features of the target image and of the plurality of reference images; and

designating the reference image corresponding to the deep features with the greatest degree of matching as the reference image matching the target image.