Sector-ring HOG for rotation-invariant human detection

Abstract

Automatic identification for victims using unmanned platforms is essential to increasing the efficiency of post-disaster searching and rescuing efforts. Different from pedestrians in upright poses, human bodies at the disaster scenes may appear in images under different planar rotations and hence rotation-invariant detection is necessary. This paper focuses on the computer vision and proposes a new visual feature, i.e. Sector-ring HOG (SRHOG), to facilitate the rotated human detection. The new feature is achieved by modifying the gradient binning and spatial binning based on the histograms of oriented gradients (HOG) and can express the rotation of image patch into the cyclic shift of final descriptor. Employing it in rotated human detection can avoid the interpolation approximation existing in image rotation and increase the detection performance. Experimentally, we first confirm the sufficient discrimination power of SRHOG in a public pedestrian dataset and further demonstrate the feasibility of bigger detection windows in an extension of this pedestrian dataset. Finally, the high performance of SRHOG for rotated human bodies is demonstrated on a victim dataset which has varied backgrounds, body postures and shear changes.