Pedestrian detection 2 (Development History of Pedestrian detection)

Pedestrian detection plays a vital role in many applications in the computer vision field, such as video surveillance, vehicle driver assistance systems, and human motion capturing systems. image Pedestrian detection methods can be divided into two categories: contour matching and apparent features. the apparent feature is defined as the image feature space (also called the descriptive operator). It can be divided into the integral method, the local method, and the feature point-to-point method.

In the overall method, Papageorgiou and Poggio [1] proposed the Haar wavelet (HWS) features and used SVM to train pedestrians, including the front and back of pedestrians. viola and Jones [2, 3] use the cascade AdaBoost learning algorithm (that is, a strong classifier is selected based on a weak classifier that exceeds a certain threshold) extract basic Haar-like features and extended Haar-like features for pedestrian detection in video surveillance. levi and Weiss [4] propose Edge Orientation histograms (eohs) for face detection. eohs first calculates the gradient intensity of the image, and then divides the imageKAnd Feature Representation is a real value obtained through the statistical ratio between directions. Both Haar-like and eohs can speed up the operation through the image integral graph method.

Dalal and triggs [5] proposed histogram of Oriented Gradient (hog ). based on the gradient information, hog allows overlapping blocks. Therefore, it is not sensitive to illumination changes and offsets and can effectively portray the edge features of the human body. however, hog also has its disadvantages: High feature dimensions, a large number of overlapping and histogram statistics, making feature computing slow, thus affecting real-time performance; poor occlusion processing capability; color, shape, texture, and other features are not used. in view of these shortcomings, some researchers have put forward more pedestrian characteristics of [6, 7, 8, 9], including cov, integral channel feature, ACF, and GGP.

Chen et al. [10] proposed Weber Local Descriptor (wld ). wld is composed of two parts: Differential excitation and orientation. wld makes full use of the Weber theorem of the human visual mechanism to have a certain degree of robustness to light and shade changes and noise interference, the disadvantage is that the computation is complicated.

The main idea of the local method is to regard the human body as a combination of parts. This method should solve two problems: Construct an effective location detector and the ry relationship between the modeling parts. mohan and others [11] divide the human body into four parts: Head, lower body, and left and right arm, and take the response values of classifier in each part as the input of SVM, construct a combined multi-level classifier to detect pedestrians.

Edgelet features describe the contour characteristics of the human body, but they describe the characteristics of the local contour of the human body, including shapes such as straight lines and arcs. it divides the human body into several parts for training, such as the whole body, Head, shoulder, leg, and trunk. Each part uses the Adaboost algorithm to train a strong classifier, use the joint probability of four parts for decision-making. this algorithm uses local features of the human body, so it still performs well in case of occlusion. The disadvantage is that feature calculation is complicated.

Wu [12] edgelet features extracted from images are used to detect the human body in static images. models are created for each part of the human body. Each edgelet describes the outlines of a specific part of the human body, then, the Adaboost algorithm is used to screen out the most effective edgelet group to describe the whole person.

Wu defines three types of edgelet, including linear, arc, and symmetric. each edgelet consists of a group of edge points. It is a line segment with a certain shape and position. for any position in the image, a response value is obtained based on whether the position has an edge similar to an edgelet shape. if the edge shape is similar to edgelet, the response value is higher.

These Methods Detect the Partial Area of the window respectively, and then combine the detection results of these areas to make the final decision. the advantage is that it can better deal with occlusion and the diversity of pedestrian posture. The main problem is how to define the local area and how to integrate information from multiple location detectors.

The feature point-to-point method regards pedestrian detection as a generalized Hof Transformation: first, the local feature detector is used to find the key points. Then, a fixed size image block is selected around the key points, the spatial distribution mode of image blocks is established by means of clustering, random forest, or maximum interval. Finally, the pedestrian location in the image is searched through the HOV voting method. A typical method is the sift feature proposed by David Lowe [13].

Whether it is the integral, local, or feature point-to-point method, the core issue is how to effectively express the overall features, regional features, or local block features of pedestrians. the histogram feature of the direction gradient is a widely used pedestrian Feature Representation. However, the direction gradient cannot depict human visual sensitivity, and the information redundancy is high. in this paper, based on the center transform histogram [14] feature (census transform histogram, centrist for short), a distinctive texture structure feature is proposed, unlike the centrist histogram, which is similar to a Local Binary Pattern histogram, the feature integrates the laws of human visual psychology to better detect pedestrians in road environments such as illumination fluctuations and background clutter.

[1] Papageorgiou C, Poggio t. a trainable System for object detection. Int 'l Journal of computer vision, (1): 15-33.

[2] Viola P, Jones MJ, snow D. detecting pedestrians using patterns of motion and appearance. in: Proc. of the Int 'l Conf. on computer vision. 2003. 734-741.

[3] Jones MJ, snow D. Pedestrian detection using boosted features over parse frames. In: Proc. Of the IEEE Conf. Computer Vision and pattern recognition. 2008. 1-4.

[4] Levi K, Weiss y. learning Object Detection from a small number of examples: the importance of good features. in: Proc. of the IEEE Conf. on computer vision and pattern recognition. 2004.53-60.

[5] Dalal N, triggs B. histograms of Oriented gradients for human detection. In: Proc. Of the IEEE Conf. on computer vision and pattern recognition. 2005.886-893.

[6] tuzel o, porikli F, Meer P. Pedestrian detection via Classification on Riemannian manifolds. IEEE Trans. On PAMI, 1713 (10): 1727.

[7] dollar P, tu Z, perona P, belongie S. Integral channel features. In: Proc. Of the British machine vision Conf. 2009. 1-11.

[8] Gao W, ai h, Lao S. adaptive contour features in oriented granular space for human detection and segmentation. in: Proc. of the IEEE Conf. on computer vision and pattern recognition. 2009. 1786-1793.

[9] Liu YZ, Shan SG, Zhang WC, Chen XL, Gao W. granularity-tunable gradients partition (GGP) descriptors for human detection. in: Proc. of the IEEE Conf. on computer vision and pattern recognition. 2009.1255-1262.

[10] Chen J, Shan SG, he c, Zhao Gy. wld: a robust Local Image Descriptor. IEEE Trans. on Pattern Analysis and machine intelligence, 2010, 32 (9): 1705-1720.

[11] Mohan A, Papageorgiou C, Poggio T. Example-based object detection in images by components. IEEE Trans. On Pattern Analysis and machine intelligence, 2001,23 (4): 349-361

[12] Wu B, nevatia R, Li Y. segmentation of multiple, partially occluded objects by grouping, merging, assigning part detection responses. int 'l Journal of computer vision, 2009,82: 185-204.

[13] Lowe DG. distinctive image features from scale-invariant keypoints. Int 'l Journal of computer vision, sp_60 (2): 91-l10.

[14] Wu JX, rehg JM. centrist: A visual descriptor for scene categorization. IEEE Trans. On Pattern Analysis and machine intelligence, 2011,33 (8): 1489-1501.

Pedestrian detection 2 (Development History of Pedestrian detection)