Improving pedestrian detection with selective gradient self-similarity feature. Issue 8 (August 2015)

• Record Type: Journal Article
• Authors: Wu, Si; Laganière, Robert; Payeur, Pierre
• Is Part Of: Pattern recognition. Volume 48:Issue 8(2015:Aug.)
• Page Start: 2364
• Page End: 2376
• Publication Date: 2015-08
• Subjects: Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2015.01.005
• Languages: English
• ISSNs: 0031-3203

Abstract

<abstract abstract-type="author" id="ab0005"> <title id="sect0005">Abstract</title> <sec> <p id="sp0090">Gradient features play important roles for the problem of pedestrian detection, especially the histogram of oriented gradients (HOG) feature. To improve detection accuracy in terms of feature extraction, HOG has been combined with multiple kinds of low-level features. However, it is still possible to exploit further discriminative information from the classical HOG feature. Inspired by the symmetrical characteristic of pedestrian appearance, we present a novel feature of gradient self-similarity (GSS) in this work. GSS is computed from HOG, and is applied to capturing the patterns of pairwise similarities of local gradient patches. Furthermore, a supervised feature selection approach is employed to remove the non-informative pairs. As a result, the selective GSS (SGSS) feature is built on a concise subset of pair comparisons. The experimental results demonstrate that significant improvement is achieved by incorporating HOG with GSS/SGSS. In addition, considering that HOG is a prerequisite for GSS/SGSS, it is intuitional to develop a two-level cascade of classifiers for obtaining improved detection performance. Specifically, the first level is a linear SVM with the multiscale HOG features to efficiently remove easy negatives. At the second stage, the already computed HOG features are reused to produce the corresponding GSS/SGSS features, and then the combined features are used to discriminate true positives from candidate image regions. Although simple, this model is competitive with the state-of-the-art methods on the well-known datasets.</p> </sec> </abstract>