Latent SVM 1 st Frame manually select target

Latent SVM 1 st Frame: manually select target Find 6 highest weighted areas in template Area of 16 blocks Train 6 SVMs on those areas Train 1 SVM on entire template 7 SVMs total

Latent SVM 2 nd Frame: Scan neighborhood with all 7 templates Get a combined score map. Find best detection.

Latent SVM - Retraining Detection is found in 2 nd frame Now each classifier is retrained. Entire frame is scanned and a new score map is created. Top 100 detections from the frame, outside of the neighborhood are used as negative samples. This scanning of the entire frame to find new negative samples is only done every 5 frames. Assumption: the scan will most likely find the same false positives. No need to rescan each frame.

Latent SVM - Retraining In each frame, a new positive sample is found from the combined score maps. To retrain the 6 sub SVMs, each of the 6 sub templates are scanned over the new positive sample to create 6 score maps. The best detection in each score map is then used for the new location of each box respectively. This allows the boxes (or parts) to move around inside of the template.

Latent SVM - Retraining Each of the 6 sub SVMs are retrained using all past positive samples and the negative samples collected for the full template. Things We Tried. . . We tried only retraining when there was a bad detection. We tried not changing the positions of the boxes until there was a bad detection. We tried only using color features for the sub templates.

This Week Finished the Latent SVM Framework. Finalized Parameters Ran Tests Compared Results to both the Blended Template Framework and the original Tracking By Detection Framework using HOG, LBP and Color features.

Results

Results

Results Framewor k HOG-LBP-Color Blended Template Latent SVM Total Frames Person 1 32 34 83 84 85 Person 2 14 85 90 15 90 Person 3 10 12 33 Person 4 72 61 90 89 90 Person 5 42 53 55 55 55 33