Context based position estimation of target of interest in videos

A system (100), comprising:

• a memory (102) storing instructions;
• one or more communication interfaces (106);
• and
• one or more hardware processors (104) coupled to the memory (102) via the one or more communication interfaces (106), wherein the one or more hardware processors (104) are configured by the instructions to:
• receive, at a two-class neural network (202), a target of interest from a frame of a video to generate a score and a class label, wherein the score and a class label are generated using a classifier of the two-class neural network, the classifier comprising a first class, and a second class, wherein the first class is specific to one or more targets and the second class is specific to one or more non-targets;
• receive, at a first part (302) of the neural network (202), a subsequent frame of the video to generate one or more convolutional feature maps;
• compute, by a cost function (308), a cost value based on the class label and the score, wherein the cost value is computed at the first part (302) of the neural network, a second part (304) of the neural network, and a softmax layer (306) of the neural network, and wherein the cost value is associated with each of the generated one or more convolutional feature maps;
• compute one or more gradient maps based on the score, the class label specific to the frame and the generated one or more convolutional feature maps;
• generate, via a target activation map generator (310) one or more target activation maps, using the one or more gradient maps and the generated one or more convolutional feature maps;
• identify, a tentative location of the target of interest based on the one or more target activation maps, and context associated thereof;
• and
• estimate, by using a peak location estimator (312), a position of the target of interest in the video using the tentative location and the identified context, wherein one or more parameters of the neural network are continually updated upon processing each frame from the video to determine one or more variations specific to the target of interest, and wherein the one or more parameters are updated based on the one or more targets and the one or more non-targets being identified in each frame of the video.

One or more non-transitory machine readable information storage mediums comprising one or more instructions which when executed by one or more hardware processors cause:

• receiving, at a two-class neural network (202), a target of interest from a frame of a video to generate a score and a class label, wherein the score and class label are generated using a classifier of the two-class neural network, the classifier comprising a first class, and a second class, wherein the first class is specific to one or more targets and the second class is specific to one or more non-targets;
• receiving, at a first part of the two-class neural network, a subsequent frame of the video to generate one or more convolutional feature maps;
• computing, by a cost function, a cost value based on the class label and the score, wherein the cost value is computed at the first part of the neural network, a second part of the neural network, and a softmax layer of the neural network, and wherein the cost value is associated with each of the generated one or more convolutional feature maps;
• computing one or more gradient maps based on the score, the class label specific to the frame and the generated one or more convolutional feature maps, wherein the one or more gradient maps are computed for each frame until a last frame of the video using the class label specific to the frame that is computed using the two-class neural network;
• generating, one or more target activation maps, using the one or more gradient maps and the generated one or more convolutional feature maps;
• identifying, a tentative location of the target of interest based on the one or more target activation maps, and context associated thereof;
• and
• estimating, by using a peak location estimator, a position of the target of interest in the video using the tentative location and the identified context, wherein one or more parameters of the neural network are continually updated upon processing each frame from the video to determine one or more variations specific to the target of interest, and wherein the one or more parameters are updated based on the one or more targets and the one or more non-targets being identified in each frame of the video.