Human detection in fish-eye images using HOG-based detectors over rotated windows

Fish-eye cameras are efficient means to provide an omni-view video recording over a large area using a single camera. Although effective algorithms for human detection in images captured by conventional cameras have been developed, human detection in fish-eye images remains an open challenge. Recognizing that humans typically appear on radial lines emitted from the center in fish-eye images, we propose to apply the popular human detection algorithm based on the Histogram of Oriented Gradient (HOG) features after rotating each search window on a radial line to the vertical reference line. We extract positive and negative examples by such rotations to train the SVM classifier using HOG features. To detect humans in a given image, we rotate the image successively and detect windows containing humans along the reference line after each rotation using the trained classifier. We use multiple window sizes to detect people with different appearance sizes. We further develop an algorithm to discover multiple overlapping windows covering the same person and identify the window that encloses the person the best. The proposed method has yielded highly accurate human detection in low-resolution, low-contrast images containing multiple people with varying poses and sizes.

1. Introduction

Human detection in video footage is an important task in many applications, including video surveillance in dynamic scenes, driving assistance system, content-based retrieval, etc. Effective algorithms have been developed for human detection in video captured by conventional cameras [1]. For video surveillance, fish-eye cameras are often used because they can cover a large region using a single camera. Because of the special characteristics of video frames captured by fish-eye cameras, human detection in fish-eye video remains an open challenge. Some previous works [2] [3] depend on first knowing intrinsic or extrinsic parameters of the fish-eye camera, then using these parameters to support their detection process. Satio et al. use the geometric relations to calculate the rough height of human in certain place [4] and use this information to guide human detection. Their algorithm is further limited to the application where people only walk through certain regions in the surveyed area. Other approaches first warp the fisheye view to a normal view and then apply human detection algorithms for normal views. This approach suffers from the inaccuracy in camera calibration and the distortion from the warping process.