Deep cross-correlation learning for object tracking

A method for tracking a target object via an artificial neural network (ANN), comprising:

• receiving a target object image of the target object extracted from a first frame of a sequence of frames, the target object being an object that is to be located in a second frame of the sequence of frames;
• receiving a search region image of a search region extracted from the second frame of the sequence of frames, the search region corresponding to an expected location of the target object in the second frame of the sequence of frames based on a location of the target object in the first frame, the first frame and the second frame being consecutive frames in the sequence of frames, the first frame and the second frame corresponding to different images, and a size of the search region image being greater than a size of the target object image;
• extracting a target region feature map of the target object image and a search region feature map of the search region image;
• receiving the extracted target region feature map and the extracted search region feature map;
• convolving the extracted target region feature map with the extracted search region feature map to determine a cross-correlation map;
• predicting coordinates of the target object in the second frame based on the cross-correlation map;
• comparing the cross-correlation map with a ground truth cross-correlation map to determine a loss value, the ground truth cross-correlation map based on a ground truth target feature map of the target object from the second frame;
• and
• back propagating the loss value into an artificial neural network to update a plurality of filter weights of a representation function.

An apparatus, comprising:

• means for receiving a target object image of a target object extracted from a first frame of a sequence of frames, the target object being an object that is to be located in a second frame of the sequence of frames;
• means for receiving a search region image of a search region extracted from the second frame of the sequence of frames, the search region corresponding to an expected location of the target object in the second frame of the sequence of frames based on a location of the target object in the first frame, the first frame and the second frame being consecutive frames in the sequence of frames, the first frame and the second frame corresponding to different images, and a size of the search region image being greater than a size of the target object image;
• means for extracting a target region feature map of the target object image and a search region feature map of the search region image;
• means for receiving the extracted target region feature map and the extracted search region feature map;
• means for convolving the extracted target region feature map with the extracted search region feature map to determine a cross-correlation map;
• means for predicting coordinates of the target object in the second frame based on the cross-correlation map;
• means for comparing the cross-correlation map with a ground truth cross-correlation map to determine a loss value, the ground truth cross-correlation map based on a ground truth target feature map of the target object from the second frame;
• and
• means for back propagating the loss value into an artificial neural network to update a plurality of filter weights of a representation function.

A non-transitory computer-readable medium having program code recorded thereon, the program code being executed by a processor of a neural computing device and comprising:

• program code to receive a target object image of a target object extracted from a first frame of a sequence of frames, the target object being an object that is to be located in a second frame of the sequence of frames;
• program code to receive a search region image of a search region extracted from the second frame of the sequence of frames, the search region corresponding to an expected location of the target object in the second frame of the sequence of frames based on a location of the target object in the first frame, the first frame and the second frame being consecutive frames in the sequence of frames, the first frame and the second frame corresponding to different images, and a size of the search region image being greater than a size of the target object image;
• program code to extract a target region feature map of the target object image and a search region feature map of the search region image;
• program code to receive the extracted target region feature map and the extracted search region feature map;
• program code to convolve the extracted target region feature map with the extracted search region feature map to determine a cross-correlation map;
• program code to predict coordinates of the target object in the second frame based on the cross-correlation map;
• program code to compare the cross-correlation map with a ground truth cross-correlation map to determine a loss value, the ground truth cross-correlation map based on a ground truth target feature map of the target object from the second frame;
• and
• program code to back propagate the loss value into an artificial neural network to update a plurality of filter weights of a representation function.