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 stream of images comprising:
a) passing an initial frame of said stream of images through a first neural network comprising a plurality of convolutional layers, each convolutional layer producing at least one feature map;
b) providing a feature map output from said first neural network to a second multi-classifier neural network comprising at least one fully-connected layer to determine a region of interest (ROI) bounding an object of a given class in said initial frame;
c) storing a feature map output from a first convolutional layer of said first neural network corresponding to said ROI as weights for at least a first layer of neurons of a third multi-layer neural network;
d) acquiring a subsequent frame from said stream of images;
e) scanning at least a portion of said frame ROI by ROI using said first neural network to produce respective feature maps from said first convolutional layer for each ROI;
f) providing said feature maps to said third multi-layer neural network to provide an output proportional to the level of match between the feature map values used for the weights of the first layer of neurons and the feature map values provided for a candidate ROI to identify a candidate ROI having a feature map best matching the stored feature map;
g) responsive to said match meeting a threshold, updating the stored feature map indicative of the features of said object according to the feature map for the best matching candidate ROI including updating the weights for said neurons according to the feature map for the best matching candidate ROI from said subsequent frame; and
h) repeating steps d) to g) until said match fails to meet said threshold.
2. A method according to claim 1 wherein each of said first and second neural networks are trained offline.
3. A method according to claim 1 wherein at least one layer of said third neural network comprises a plurality of neurons performing a sigmoid activation function on the outputs of neurons from said first layer.
4. A method according to claim 3 wherein a third layer of said third neural network comprises at least a neuron aggregating the outputs of the neurons of a second layer.
5. A method according to claim 1 wherein said neural network is a forward connected neural network.
6. A method according to claim 3 comprising providing a piecewise linear approximation of said sigmoid function.
7. A method according to claim 1 wherein said scanning comprises scanning across a scan area of a subsequent frame, ROI-by-ROI, said scan area being a function of said best matching candidate ROI from a previous frame and inter-frame movement between a previous frame and said subsequent frame.
8. A method according to claim 1 further comprising periodically performing steps a) to c).
9. A method according to claim 1 comprising only performing steps a) to c) in response to said best match not exceeding said threshold.
10. A method according to claim 1 wherein said first convolutional layer is an intermediate layer of said first neural network.
11. An image processing system including digital logic for implementing the method of claim 1 .