Dimensional Reduction with PCA for Feature Selection in Pedestrian Detection

The feature orientation gradient histogram (HOG) together with the classification support vector engine (SVM) is still considered the standard for detecting modern pedestrians using various features and classification schemes. In this study, we investigated and developed a HOG solution combined with SVM to achieve a robust detection method. With minor changes in shape, image direction, and location, as well as significant color changes, the proposed HOG descriptor performs well. While the HOG performed similarly in the other classes, we concentrated on upright people, or pedestrians, in this study. However, when using different cell sizes, overlaps, and block sizes, the HOG method will vary, resulting in the occurrence of high-dimensional feature vectors. Principal Component Analysis (PCA) is used to solve this problem by reducing the HOG dimensions from 3780 to 937. As a result, SVM can handle shorter feature vectors, requiring less computational time and effort. Following that, we experimented with various SVM models to find the best SVM parameters and kernel for the INRIA data set. Like the application of PCA on the Linear HOG-SVM detector, it produces a precision value of 0.952, a recall of 0.982, and an F-score of 0.967. Therefore, this study not only proposes a linear SVM detector combined with HOG as a reference for detecting pedestrians but also contributes to the research of similar tasks involving other features and objects..