Real-time pedestrian detection based on resolution aware feature transformation

We describe a real-time pedestrian detection system intended for use in automotive applications. Our system demonstrates competitive detection performance when compared to the state-of-the-art detectors in Far scale scenes and is able to run at a speed of 15 fps on an Intel Core i5 computer when applied to 640x480 images. To gain this achievement, we make efforts on both efficiency and effectiveness. First, we analyze geometric constraints at different resolutions in order to efficiently search the feature pyramids. Second, we use resolution aware transformation of feature subspace to increase the detection accuracy of low resolution pedestrians. The transformation is used to map features of different resolution pedestrians to a common subspace where the detector model will be trained. For model learning, we present a modification of SVM training process to learn the resolution aware transformations and SVM model based detector iteratively. We have evaluated our system on our ground truth datasets. Our system shows 0.42 miss rate with 1 false positive per image (FPPI) while its effective detection distance maintains 40m in our ground truth test. Copyright © 2014 Binary Information Press.