Object detection method and object detection apparatus

Abstract

This invention provides an object detection method and an object detection apparatus. In the method, the size of an original image is reduced to generate a reduced-size image. Object detection is performed on the reduced-size image to generate first object detection information. The original image is cropped using the first object detection information to output a key image region. Object detection is performed on the key image region to generate second object detection information. This effectively improves the detection accuracy of small or distant objects while maintaining computational efficiency.

Claims

1. An object detection method, implemented through a processor, the object detection method comprising: The image is reduced in size to produce a scaled-down image, wherein the total number of images in the scaled-down image is less than the total number of images in the original image; an object detection is performed on the scaled-down image to generate first object detection information, wherein the first object detection information includes a confidence score of the object detection for at least one detected first object, and the object detection on the scaled-down image includes: obtaining at least one first candidate object among the at least one first object having a confidence score less than a confidence threshold value; cropping the original image using the first object detection information and outputting at least one key image region in the original image corresponding to the first object detection information, wherein the total number of images in the at least one key image region is less than the total number of images in the original image, and the position of the at least one key image region corresponds to the position of at least one of the at least one first candidate object; and object detection is performed on the at least one key image region to generate second object detection information.

2. The object detection method as described in claim 1, wherein performing the object detection on the scaled-down image includes: The at least one key image region is determined based on the positional distribution of the at least one first candidate object, wherein the positional distribution includes the number of objects per unit area.

3. The object detection method as described in claim 2, wherein determining the at least one key image region based on the positional distribution of the at least one candidate object includes: Determine the number of objects in the at least one first candidate object in the region of interest corresponding to each of the at least one first candidate object, wherein the unit region is the region of interest; And by using the number of objects in the region of interest corresponding to the at least one first candidate object, at least one of the regions of interest corresponding to the at least one first candidate object is used as the at least one key image region.

4. The object detection method as claimed in claim 3, wherein identifying at least one of the regions of interest corresponding to the at least one first candidate object as the at least one key image region by using the number of objects in the region of interest corresponding to the at least one first candidate object includes: The region of interest corresponding to the at least one first candidate object is selected as the at least one key image region, which has a large number of objects.

5. The object detection method as described in claim 4, wherein the at least one key image region includes multiple key image regions, and the object detection method further includes: If there are identical first candidate objects in the key image regions, only one of them in the key image regions is retained, and the other key image regions that are not retained are removed.

6. The object detection method as claimed in claim 3, wherein determining the number of objects in the region of interest corresponding to each of the at least one first candidate object includes: The representative position of each of the at least one first candidate object is taken as the center of the corresponding region of interest, wherein the regions of interest corresponding to the at least one first candidate object are of the same size. And count the first candidate objects covered by the region of interest corresponding to each of the at least one first candidate object, to generate the number of objects in the region of interest corresponding to the at least one first candidate object, wherein the proportion of the area covered by the corresponding region of interest for the counted first candidate objects is greater than the coverage threshold value.

7. The object detection method as described in claim 1 further includes: The first object detection information and the second object detection information are integrated to generate a final object detection information, wherein the final object detection information includes at least one first object corresponding to both the first object detection information and the second object detection information.

8. The object detection method as claimed in claim 7, wherein the second object detection information includes a confidence score of the object detection for at least one detected second object, and integrating the first object detection information and the second object detection information includes: Obtain at least one second candidate object from the at least one first object that has a confidence score not less than the confidence threshold value; Obtain at least one third candidate object from the at least one second object that has a confidence score not less than the confidence threshold value; and determine the final object detection information based on an intersection area between the at least one second candidate object and the at least one third candidate object, wherein the intersection area corresponds to the portion of the bounding box corresponding to the at least one second candidate object and the bounding box corresponding to the at least one third candidate object that overlaps.

9. The object detection method as described in claim 8, wherein determining the final object detection information based on the intersection area of ​​the at least one second candidate object and the at least one third candidate object includes: Determine the first intersection ratio of the intersection area to the bounding box corresponding to the at least one second candidate object; Determine a second intersection ratio of the intersection area to the bounding box corresponding to the at least one third candidate object; and remove the at least one third candidate object based on the first intersection ratio and the second intersection ratio, wherein the final object detection information excludes the removed third candidate object.

10. The object detection method as claimed in claim 1, wherein the second object detection information includes a confidence score of the object detection for at least one detected second object, and the object detection method further includes: Obtain at least one second candidate object from the at least one first object that has a confidence score not less than the confidence threshold value; Obtain at least one third candidate object from the at least one second object that has a confidence score not less than the confidence threshold value; And based on the increase in the number of the at least one second candidate object in each of the at least one key image regions exceeding the number of the at least one third candidate object, at least one of the at least one key image regions is selected as at least one anchored image region, wherein the at least one anchored image region is used for object detection of the original image of the next frame.

11. The object detection method as described in claim 10 further includes: The position of the corresponding anchored image area is changed based on the representative position of the third candidate object in the selected key image area.

12. An object detection device, comprising: A memory module used to store a piece of code; and a processor, coupled to the memory and configured to load the code to execute: reducing the size of an original image to produce a scaled-down image, wherein the total number of images in the scaled-down image is less than the total number of images in the original image; performing object detection on the scaled-down image to generate first object detection information, wherein the first object detection information includes a confidence score of the object detection for at least one detected first object, and the processor further performs: obtaining at least one first candidate object among the at least one first object having a confidence score less than a confidence threshold; cropping the original image using the first object detection information and outputting at least one key image region in the original image corresponding to the first object detection information, wherein the total number of images in the at least one key image region is less than the total number of images in the original image, and the position of the at least one key image region corresponds to the position of at least one of the at least one first candidate object; and performing object detection on the at least one key image region to generate second object detection information.

13. The object detection apparatus of claim 12, wherein the processor further performs: determining the at least one key image region based on the positional distribution of the at least one first candidate object, wherein the positional distribution includes the number of objects per unit area.

14. The object detection apparatus of claim 13, wherein the processor further performs: determining the number of objects in a region of interest corresponding to each of the at least one first candidate object, wherein the unit region is the region of interest; and using the number of objects in the region of interest corresponding to the at least one first candidate object as the at least one key image region.

15. The object detection apparatus of claim 14, wherein the processor further performs: finding at least one of the regions of interest corresponding to the at least one first candidate object that has a larger number of objects as the at least one key image region.

16. The object detection apparatus of claim 15, wherein the at least one key image region comprises a plurality of key image regions, and the processor further performs: in response to having the same first candidate object in the key image regions, retaining only one of the key image regions and removing the other key image regions that are not retained.

17. The object detection apparatus of claim 14, wherein the processor further performs: taking a representative position of each of the at least one first candidate object as the center of a corresponding region of interest, wherein the regions of interest corresponding to the at least one first candidate object are of the same size; and counting the first candidate objects covered by each region of interest corresponding to the at least one first candidate object to generate the number of objects in the region of interest corresponding to the at least one first candidate object, wherein the proportion of the area covered by the corresponding region of interest of the counted first candidate objects is greater than a coverage threshold value.

18. The object detection apparatus of claim 12, wherein the processor further performs: integrating the first object detection information and the second object detection information to generate final object detection information, wherein the final object detection information includes at least one first object corresponding to both the first object detection information and the second object detection information.

19. The object detection apparatus of claim 18, wherein the second object detection information includes a confidence score of the object detection for at least one detected second object, and the processor further performs: obtaining at least one second candidate object among the at least one first object having a confidence score not less than the confidence threshold value; obtaining at least one third candidate object among the at least one second object having a confidence score not less than the confidence threshold value; and determining the final object detection information based on an intersection area of ​​the at least one second candidate object and the at least one third candidate object, wherein the intersection area corresponds to the portion of the bounding box corresponding to the at least one second candidate object overlapping the bounding box corresponding to the at least one third candidate object.

20. The object detection apparatus of claim 19, wherein the processor further performs: determining a first intersection ratio of the intersection area to the bounding box corresponding to the at least one second candidate object; determining a second intersection ratio of the intersection area to the bounding box corresponding to the at least one third candidate object; and removing the at least one third candidate object based on the first intersection ratio and the second intersection ratio, wherein final object detection information excludes the removed third candidate object.

21. The object detection apparatus of claim 12, wherein the second object detection information includes a confidence score of the object detection for at least one detected second object, and the processor further performs: obtaining at least one second candidate object among the at least one first object having a confidence score not less than the confidence threshold value; obtaining at least one third candidate object among the at least one second object having a confidence score not less than the confidence threshold value; and selecting at least one from the at least one key image region as at least one anchored image region based on the increase in the number of the at least one second candidate object exceeding the at least one third candidate object in each of the at least one key image region, wherein the at least one anchored image region is used for the object detection of the original image of the next frame.

22. The object detection apparatus of claim 21, wherein the processor further performs: changing the position of the corresponding anchored image region based on the representative position of a third candidate object in the selected key image region.

Images