Method and system for tracking object

Abstract

A method of tracking an object across a stream of images comprises passing an initial frame of the stream through a first neural network comprising a plurality of convolutional layers. A feature map output from the first neural network is provided to a second multiclassifier neural network comprising at least one fully-connected layer to determine a region of interest (ROI) bounding an object of agiven class in the initial frame. A feature map output from a first convolutional layer of the first neural network corresponding to the ROI is stored as weights for at least a first layer of neuronsof a third multi-layer neural network. A subsequent frame from the stream is acquired and the frame is scanned ROI by ROI using the first neural network to produce respective feature maps from the first convolutional layer for each ROI. The feature maps are provided to the third multi-layer neural network to provide an output proportional to the level of match between the feature map values usedfor 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. The storedfeature map weights are updated with weights from the best matching candidate ROI.

Description

technical field [0001] The present invention relates to a method and system for tracking an object of interest across an image stream. Background technique [0002] Viola-Jones in the following US 2002 / 0102024 discloses a method for detecting regions of interest (ROIs) including objects such as faces within captured images, typically in video streams image frame. Briefly, Viola-Jones first derives an integral image from the acquired images. Each element of the integral image is computed as the sum of the intensities of all points in the image to the upper left of that point. The total intensity for any sub-window in the image can then be derived by subtracting the integrated image value of the upper left sub-window from the integrated image value of the lower right sub-window. The intensities of adjacent sub-windows can be efficiently compared using specific combinations of integral image values ​​of points from the sub-windows. [0003] The Viola-Jones based object dete...

AI Technical Summary

Problems solved by technology

[0004] It should be appreciated that applying Viola-Jones analysis to each part of the image for each size of object to be detected is still quite processor intensive, which may prevent the system from running fast enough to detect and track in real time objects in the image stream

Images