Training for camera lens distortion
Cameras with large field-of-view lenses can cause significant geometric distortions of the images acquired. Training for object detection normally takes place on undistorted images. Thus, in order to detect objects of interest within the acquired images, an undistortion procedure is applied on the acquired images and an object detection is then performed on the undistorted images. Unfortunately, such undistortion procedure is too computationally expensive to be run on some edge devices. To remove the need to perform the undistortion procedure, it is proposed to train for object detection directly from distorted acquired images.
1. A method of processing images, comprising:
obtaining, by an object detection device, one or more captured images by capturing one or more images through a lens of the object detection device, the one or more captured images including distortions caused by the lens;
receiving an object detection model from a training processor, the object detection model having been trained to recognize one or more objects of interest from distorted training images;
training a plurality of object classifiers of the object detection device with the object detection model; and
detecting, by the object detection device, the one or more objects of interest directly from the one or more captured images without undistorting the one or more captured images prior to performing the object detection,
wherein the detecting includes classifying the one or more objects.
2. The method of claim 1 , wherein detecting the one or more objects of interest comprises processing the one or more captured images through the plurality of object classifiers of the object detection device.
3. The method of claim 1 , wherein the plurality of object classifiers comprise a plurality of local binary pattern (LBP) classifiers.
4. The method of claim 1 , wherein detecting the one or more objects of interest comprises detecting at least one object of interest occupying a region of at least one captured image corresponding to an edge of the lens.
5. The method of claim 1 ,
wherein detecting the one or more objects of interest comprises detecting at least one object of interest objects occupying a peripheral region of at least one captured image, and
wherein the peripheral region is a region between pixels of the at least one captured image corresponding to a threshold angle of the lens and pixels of the at least one captured image corresponding to an edge of a field-of-view (FOV) of the lens, the FOV of the lens being wider than the threshold angle.
6. An object detection device comprising:
a processor;
a plurality of object classifiers; and
a memory,
wherein the processor, the plurality of object classifiers, and the memory are configured to:
obtain one or more captured images by capturing one or more images through a lens of the object detection device, the one or more captured images including distortions caused by the lens,
receive an object detection model from a training processor, the object detection model having been trained to recognize one or more objects of interest from distorted training images,
train the plurality of object classifiers with the object detection model; and
detect the one or more objects of interest directly from the one or more captured images without undistorting the one or more captured images prior to performing the object detection, and
wherein the detecting includes classifying the one or more objects.
7. The object detection device of claim 6 , wherein the plurality of object classifiers are configured to process the one or more captured images to detect the one or more objects of interest.
8. The object detection device of claim 6 , wherein the plurality of object classifiers comprise a plurality of local binary pattern (LBP) classifiers.
9. The object detection device of claim 6 , wherein the processor, the plurality of object classifiers, and the memory are configured to detect at least one object of interest occupying a region of at least one captured image corresponding to an edge of the lens.
10. The object detection device of claim 6 ,
wherein the processor, the plurality of object classifiers, and the memory are configured to detect at least one object of interest occupying a peripheral region of at least one captured image, and
wherein the peripheral region is a region between pixels of the at least one captured image corresponding to a threshold angle of the lens and pixels of the at least one captured image corresponding to an edge of a field-of-view (FOV) of the lens, the FOV of the lens being wider than the threshold angle.