Pedestrian Recognition Machine Perception and Modeling of Human Behavior Manfred Lau
Pedestrian Recognition Oren, Papageorgiou, Sinha, Osuna, Poggio. Pedestrian Detection Using Wavelet Templates. CVPR Papageorgiou, and Poggio. Trainable Pedestrian Detection. International Conference on Image Processing 1999.
Motivation Recognition system inside vehicles Valerie – detect and greet those who stop in front of the booth
Overview Positive samplesNegative samples Classifier
Wavelet Template 1 vertical wavelet Average of many samples Compute coefficient for each RGB channel and take largest absolute value Vertical wavelet identifies “vertical color differences”
Wavelet Template 1 vertical horizontal diagonal 1 1 Average of many samples
Features Each image is one instance with 1326 features and one classification Same thing for negative samples
Test case 282 positive samples, 236 negative samples for training 20 positives and 20 negatives for testing Some Positive Samples
Some negative samples
Results Nearest neighbor classifier 95% accuracy Decision tree classifier 90% accuracy 2 false positives3 false positives, 1 false negative
10-fold cross validation Test case: 302 positives, 256 negatives Nearest neighbor 94.27% 30 false positives, 2 false negatives Decision tree 86.74% 47 false positives, 27 false negatives
Incremental bootstrapping Use nearest neighbor But problem with many false positives
Incremental bootstrapping Took database of 558 total samples After bootstrapping, 656 total samples
Bootstrapping
Result A completely new test image Before bootstrapping 85.06% accurate, 65 false pos, 0 false neg After bootstrapping 90.11% accurate, 43 false pos, 0 false neg
Result Another new test image Before bootstrapping 75.86% accurate, 100 false pos, 5 false neg After bootstrapping 81.15% accurate, 77 false pos, 5 false neg
Splitted up into 560 images, about 30 classified as positive Some false positives true positives
Results
Less features Take average coefficients across many positive samples Pick those features that are darkest/lightest can use much less than 1326 features, for faster classification
Conclusions Can detect positive samples well, but many false positives Bootstrapping on more and more new images will decrease false positives (I’m not doing enough of this)
Limitations Recognize only template, other objects may be similar Difficult to define what is a negative sample What if pedestrians are partially occluded?