Method and system for tracking an object
A method of tracking an object across a stream of images comprises determining a region of interest (ROI) bounding the object in an initial frame of an image stream. A HOG map is provided for the ROI by: dividing the ROI into an array of M×N cells, each cell comprising a plurality of image pixels; and determining a HOG for each of the cells. The HOG map is stored as indicative of the features of the object. Subsequent frames are acquired from the stream of images. The frames are scanned ROI by ROI to identify a candidate ROI having a HOG map best matching the stored HOG map features. If the match meets a threshold, the stored HOG map indicative of the features of the object is updated according to the HOG map for the best matching candidate ROI.
1. A method of tracking an object across a set of image frames, the method comprising:
inputting a first frame of the set of image frames into a first neural network, the first neural network comprising at least one convolutional layer and at least one fully-connected layer;
receiving a first output from the first neural network representative of a first map value associated with the first frame and a region of interest in the first frame that comprises the object;
determining a weight value based at least in part on the first map value and the region of interest;
inputting a second frame of the set of image frames into the first neural network;
receiving a second output from the first neural network representative of a second map value associated with the second frame;
inputting the second map value and the weight value into a second neural network;
receiving a third output from the second neural network identifying a region of interest in the second frame as matching the region of interest in the first frame; and
determining a location of the object in the second frame based at least in part on the third output.
2. The method as claim 1 recites, wherein:
the first neural network comprises multiple convolutional layers, and
each convolutional layer of the multiple convolutional layers produces at least one map value.
3. The method as claim 1 recites, wherein the second neural network is a classification neural network configured to provide the third output based at least in part on a class of the object.
4. The method as claim 1 recites, further comprising:
tracking the object in a third frame after the second frame based at least in part on the third output.
5. The method as claim 1 recites, further comprising:
normalizing the first output or the weight value,
wherein receiving the second output is based at least in part on the normalizing.
6. The method as claim 1 recites, wherein at least one of the first neural network or the second neural network is trained offline.
7. The method as claim 1 recites, wherein:
the second neural network is a multi-layer neural network, and
at least one layer of the multi-layer neural network comprises a sigmoid activation function.
8. The method as claim 7 recites, wherein the at least one layer of the multi-layer neural network further comprises a piecewise linear approximation of the sigmoid activation function.
9. The method as claim 1 recites, wherein the second neural network is a forward connected neural network.
10. An image processing system arranged to track an object across a set of image frames, the image processing system comprising a processor arranged to:
input a first frame of the set of image frames into a first neural network, the first neural network comprising at least one convolutional layer and at least one fully-connected layer;
receive a first output from the first neural network representative of a first map value associated with the first frame and a region of interest in the first frame that comprises the object;
determine a weight value based at least in part on the first map value and the region of interest;
input a second frame of the set of image frames into the first neural network;
receive a second output from the first neural network representative of a second map value associated with the second frame;
input the second map value and the weight value into a second neural network;
receive a third output from the second neural network identifying a region of interest in the second frame as matching the region of interest in the first frame; and
determine a location of the object in the second frame based at least in part on the third output.
11. The image processing system as claim 10 recites, wherein the second neural network is a classification neural network configured to provide the third output based at least in part on a class of the object.
12. The image processing system as claim 10 recites, wherein the first map value comprises a histogram of gradients map value.
13. The image processing system as claim 10 recites, wherein the first map value identifies a feature of the region of interest in the first frame.
14. The image processing system as claim 10 recites, wherein at least one of the first neural network or the second neural network is trained offline.
15. The image processing system as claim 10 recites, wherein:
the second neural network is a multi-layer neural network, and
at least one layer of the multi-layer neural network comprises a sigmoid activation function.
16. The image processing system as claim 15 recites, wherein the at least one layer of the multi-layer neural network comprises a piecewise linear approximation of the sigmoid activation function.
17. The image processing system as claim 10 recites, wherein the second neural network is a forward connected neural network.
18. A device arranged to track an object across a set of image frames, the device comprising a processor arranged to:
input a first frame of the set of image frames into a first neural network, the first neural network comprising at least one convolutional layer and at least one fully-connected layer;
receive a first output from the first neural network representative of a first map value associated with the first frame and a region of interest in the first frame that comprises the object;
determine a weight value based at least in part on the first map value and the region of interest;
input a second frame of the set of image frames into the first neural network;
receive a second output from the first neural network representative of a second map value associated with the second frame;
input the second map value and the weight value into a second neural network; and
receive a third output from the second neural network identifying a region of interest in the second frame as matching the region of interest in the first frame; and
determine a location of the object in the second frame based at least in part on the third output.
19. The device as claim 18 recites, the processor further arranged to determine a location of the object in a third frame of the set of image frames based at least in part on the third output.
20. The device as claim 18 recites, wherein the weight value is a first weight value and the processor is further arranged to:
determine a second weight value based at least in part on the second map value and the region of interest in the second frame; and
store the second weight value.