Abstract
Histogram of oriented gradient (HOG) features in combination with support vector machine (SVM) classification are still widely reported as a de facto standard benchmark to evaluate contemporary pedestrian detectors employing a vast spectrum of sophisticated features and classification schemes. In this paper, however, it has been shown that these benchmark figures for the original HOG-linear SVM detector grossly underestimate its true working potential. An improved SVM training methodology has been proposed that can make this important detector yield up to 14% lower miss rates than being currently reported on standard datasets. The proposed method accomplishes this by considering only a small set of the most relevant training examples and mitigating the class imbalance problem through manipulation of misclassification cost ratios. The second contribution of this work is the investigation of quantization errors which are induced when fast look-up table (LUT)-based kernel SVM classification is used. Based on the findings, a non-linear quantization scheme for the HOG features has been proposed to facilitate fast evaluation of kernel SVM resulting in up to 4% further reduction in miss rates without incurring additional computational cost. Hence, the linear and kernel SVM-based detectors proposed in this paper not only serve as much more stringent benchmarks for pedestrian detection but also provide the practitioners with an improved training framework for similar detection tasks involving other features and objects. The retrained models have been publicly released and provide a much better replacement for those currently being shipped with OpenCV and MATLAB computer vision libraries.