Improved Detection of Large-Sized Pedestrians Using Non-linear Scale Space and Combination of HOG and Dense LDB Features

Multi-scale detection in pedestrian detection plays a vital step due to its use in detecting pedestrians in different scales. However, multi-scale detection is a challenging task due to various reasons like selection of appropriate scale factor, low resolution and loss of sharp edges at deeper pyramidal levels, etc. In this paper, new pedestrian detectors are presented which uses non-linear scale space for multi-scale detection. The proposed detector uses histogram of oriented gradient features in combination with dense local difference binary features. Different classifiers like linear SVM and cascade of boosted classifiers are used to train the detector. INRIA pedestrian dataset is used to train and test the proposed detectors. The proposed system is evaluated in terms of precision versus recall and miss-rate versus FPPW/ FPPI as well as computational speed. The performance of the proposed detectors is also compared with some similar existing pedestrian detectors.