Target detection, tracking, and classification in compressive measurement domain
The present invention is to provide a method and system using compressed data directly for target tracking and target classification in videos. The present invention uses a video imager to generate compressive measurements, and a random subsampling operator to compress the video data. It uses a Gaussian Mixture Model (GMM) for target detection and manual location of the target and putting a bounding box around the targets in the first frame is not required. It further applies a saliency-based algorithm to re-center the captured target. This re-centering process can be repeated multiple times and each application of re-centering will improve over the previous one. A pixel completion algorithm is used to fill in the missing pixels for the captured target area. A Sparse Representation Classification (SRC) for target classification. Both the target templates in a dictionary and captured targets are transformed to the frequency domain using Fast Fourier Transform (FFT).
1. A system for target detection, tracking and classification directly in a compressive measurement domain, comprising:
an Imager having a Motion Detector (MD) means connected to a Saliency Based Re-centering (SBR) generator means for outputting a target location;
a Missing Data Completion (MDC) generator connected to the output of the SBR generator means;
output of the MDC generator is connected to a Target Classification (TC) generator;
a random subsampling operator means to compress video data; and
the output of the TC generator is connected to a Voting Based Scheme (VBS) generator means for generating a target type of image.
2. A system for target detection, tracking and classification directly in a compressive measurement domain in accordance to claim 1 , further comprising:
a Fast Fourier Transform (FFT) is connected between the MDC generator and the TC generator; and
another FFT is connected between a Target Dictionary (TD) and the TC generator.
3. A system for target detection, tracking and classification directly in a compressive measurement domain in accordance to claim 1 , wherein the MD means is based on a Gaussian Mixture Model (GMM).
4. A system for target detection, tracking and classification directly in a compressive measurement domain in accordance to claim 1 , wherein the TC generator is based on a Sparse Representation Classification (SRC).
5. A system for target detection, tracking and classification directly in a compressive measurement domain in accordance to claim 4 , wherein the TC is using a subset of pixels in each frame.
6. A method of target detection, tracking and classification directly in a compressive measurement domain, comprising the steps of:
automatic detecting a target location in a first frame using a Gaussian Mixture Model (GMM);
re-centering the target location using a Saliency-based detection;
filling in missing pixels in the detected target;
classifying the target using a frequency domain Sparse Representation Classification (SRC) approach; and
further enhancing the target classification performance using compressed data directly.
7. A method of target detection, tracking and classification directly in a compressive measurement domain in accordance to claim 6 , further comprising the steps of:
saving data storage space by using the compressed data; and
reducing the transmission bandwidth of videos.
8. A method of target detection, tracking and classification directly in a compressive measurement domain in accordance to claim 6 , further comprising the steps of:
repeating the re-centering process multiple times to improve accuracy of the target location.
9. A method of target detection, tracking and classification directly in a compressive measurement domain in accordance to claim 6 , wherein
a pixel completion algorithm is used to fill in the missing pixels for the captured target location.
10. A method of target detection, tracking and classification directly in a compressive measurement domain in accordance to claim 6 , further comprising the steps of:
transforming target templates in a dictionary and the detected targets to a frequency domain using Fast Fourier Transform (FFT).
11. A method of target detection, tracking and classification directly in a compressive measurement domain in accordance to claim 6 , further comprising the steps of:
applying a decision level fusion based on voting to further improve the classification performance.