Human Area Refinement for Human Detection

Abstract

Abstract. Human detection technologies are very useful tools to understand human activity for various purposes, such as surveillance. Recently, trackingby-detection methods have also become popular for analyzing human activity, but their performance is greatly affected by the accuracy of detected human areas because they use online learning based on the detected results. In order to improve the performance of such tracking methods, the inclination of human bodies in the image is considered as a way to refine the detected human bounding boxes. Based on background subtraction and a novel scheme of estimating human foot position, a refinement scheme is proposed to estimate a bounding box more accurately, which can better fit the contours of inclined human bodies than the conventional method. Experimental results illustrated that the bounding boxes refined by the proposed algorithm achieved a higher cover rate of 92.7 % and a smaller mean angle error of 0.7° compared with the cover rate of 83.7 % and mean angle error of 3.8° obtained using the conventional method, as determined by comparison with the ground truth, and a real-time detection speed of 32.3 fps on a 640 × 480 video has been realized. Thus, tracking performance is significantly enhanced by refining the human areas, with a mean improvement of 42.4 % in the F-measure when compared with the conventional method. [3], and the CENTRIST feature [4]. The effectiveness of these methods has been proven in practice for the detection of upright complete humans. With the development of human detection technologies, an approach called tracking-by-detection [5] has become popular recently. This approach treats the tracking problem as a detection task applied over time. Such a method learns classifiers for tracking online using detected Human Area Refinement for Human Detection 131 human bounding boxes (b-boxes) instead of using offline labeled data for training, and thus the quality of the classifiers is greatly affected by the accuracy of the detected human areas, which contributes to the final tracking performance.Although most detection methods can provide a high detection rate, accurate depiction of human postures and regions still cannot be achieved, i.e., all existing methods can only detect approximate human locations denoted by upright b-boxes, and cannot deal with the contour of an inclined human body very well. In order to improve the accuracy of the detected human areas, in this paper we propose a refinement algorithm for the detected human bounding box (b-box) to fit the contour of the inclined human body based on background subtraction, human detection, and a novel scheme of estimating human head and foot position using a predefined human height.The rest of this paper is organized as follows. Section 2 briefly introduces related work. Section 3 describes the details of the proposed approach. Section 4 presents the experimental results and discussion, and Section 5 concludes the paper. In addition, real-time detection has attracted more and more attention,...

Anthropometric Ratios for Lower-Body Detection Based on Deep Learning and Traditional Methods

“…Background subtraction [30] was introduced for object detection to identify moving objects based on the differences between the current frame and a background frame in either a pixel-by-pixel or window-by-window fashion [31]. This background subtraction was also combined the depth information to extract an object with higher success and capability [32].…”