Case Study on Histogram of Oriented Gradient (HOG) based Pedestrian Detection Methods
In a machine vision system, the critical research question that we always ask about is how machines know what machines do not know. Before doing every machine vision researches, we are always looking through the fact of what we do not know and how we know on this point.
A relevant role of features in images will be applicable for the detection of different scales pedestrians because the feature information of a large-scale image is different from a corresponding feature extracted from a small-scale image. Concerning related studies, the Histogram of Oriented Gradient [1] is a well-known feature descriptor for pedestrian detection to calculate the gradient orientation, but this technique can detect well when the full body of the pedestrian appears in the system. HOG has been widely used in pedestrian detection with the useful features of gradient and magnitude information to capture the edge and shape of the object under the variation of illumination changes. The performance of the system improved the various descriptor parameters, scale of gradients, orientation bin, coarse spatial bin, and high-quality blocks. To reduce the gradient operations in the HOG feature calculation, the fast calculation of the HOG feature [2] has been proposed to improve the detection speed of pedestrians without any loss of detection performance. A trilinear interpolation technique then applied at the HOG feature calculation step to avoid aliasing and improved detection accuracy. However, the computational complexity significantly increased due to the overlapping block-based operation.
To improve the detection performance, several features have been combined with HOG features. An automated pedestrian detection system [3] has been proposed by combining pedestrian motion patterns and HOG features in dynamic background to recognize pedestrians for reducing accidents between the vehicles and the pedestrians. Features extracted from different body portions analyzed for interpreting the presence or absence of a person in a particular region in front of the car. This feature approaches significantly improved the performance of pedestrian detection with the use of motion vectors.
Then, a fast multi-feature-based algorithm [4] has been proposed for reducing the computation complexity for pedestrian detection combined with the Histogram of Oriented Gradient (HOG) and Discrete Wavelet Transform (DWT). To speed up the pedestrian detection process, the magnitude of the motion information was used to set a region of interest (ROI) to distinguish the pedestrian candidates in the whole image for the hypothesis generation step. To reduce the computational complexity, the intensive algorithms performed pedestrians’ identification in the ROIs by extracting the multi-features between HOG and DWT for the hypothesis verification step. After that, the combination of the feature vectors was given to learn by the Support Vector Machine (SVM) mechanism to classify whether a pedestrian or not in a candidate window. The proposed algorithm achieved an increased detection rate of 27.21% by comparing it with the exciting HOG feature.
Furthermore, a co-occurrence HOG (CoHOG) algorithm was robust to illumination and poses variations for the effective classification of the rotational invariant texture to improve the HOG algorithm [5]. The co-occurrence matrix-based CoHOG feature descriptor obtained from a two-dimensional histogram pair of gradient orientations, which was an extended HOG to represent the spatial relationship between gradient orientations. An efficient global derivative scheme also proposed with the use of the Gaussian derivative filter and Hessian matrix for feature extraction. The proposed descriptor used the Gaussian curvature and mean curvature information instead of directly using the responses of gradients to develop the classical HOG and CoHOG algorithms. In this paper, four extensions of HOG and CoHOG have also been presented for efficient and reasonable use of derivative information in the local texture patterns. The promising results attained a comprehensive performance compared with the state-of-the-art texture classification methods and achieved higher classification performance with the reduction of the feature vector size of the original CoHOG algorithm.
Furthermore, the current single feature has been proved difficult to handle the occluded parts of the pedestrians, and the challenging problem among the pedestrian detection systems would hardly detect/classify the partially occluded part of the pedestrian’s body under the complex scenes. Partial occlusion mostly happened when the pedestrian will partly cover with an object and likely occur when the pedestrian was not in full view of the driver. Moreover, the occlusion and deformation can vary on different visible human body parts under complex backgrounds. A well-known method, the Deformable Part-based detectors, proposed the whole body part into multiple parts of the specific structure of the human body to handle the occlusion problem. In the past decade, many evaluations have been done to improve occluded pedestrian detection [6]. The mixtures of multiscale deformable part models introduced the context of the multiple components (i.e., head, right shoulder, left shoulder, upper leg, and lower leg) into the detector. This detector relied mainly on the discriminative training that used latent information for matching deformable models to images as well. Compared with the state-of-the-art results, the proposed model achieved efficient and accurate performance on difficult datasets.
In contrast, learning from each specific part of the human body’s structure and combining it with different features to make use of the contextual information are commonly used among occluded pedestrian handling. Most of these approaches attained the promising performance to handle the misclassification of the occluded pedestrians and estimated the location of these parts in the detection window.
References
[1] N. Dalal and B. Triggs, “Histograms of oriented gradients for human detection,” in Proceedings — 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, CVPR 2005, 2005.
[2] S. Kim and K. Cho, “Fast calculation of histogram of oriented gradient feature by removing redundancy in overlapping block,” J. Inf. Sci. Eng., vol. 30, no. 6, pp. 1719–1731, 2014.
[3] H. Ramzan, B. Fatima, A. R. Shahid, S. Ziauddin, and A. A. Safi, “Intelligent Pedestrian Detection using Optical Flow and HOG,” 2016.
[4] G. S. Hong, B. G. Kim, Y. S. Hwang, and K. K. Kwon, “Fast multi-feature pedestrian detection algorithm based on histogram of oriented gradient using discrete wavelet transform,” Multimed. Tools Appl., 2016.
[5] K. Hanbay, N. Alpaslan, M. F. Talu, D. Hanbay, A. Karci, and A. F. Kocamaz, “Continuous rotation invariant features for gradient-based texture classification,” Comput. Vis. Image Underst., vol. 132, pp. 87–101, 2015.
[6] P. F. Felzenszwalb, R. B. Girshick, D. Mcallester, and D. Ramanan, “Object Detection with Discriminatively Trained Part-Based Models.”
