Rotation-Invariant Human Detection Scheme Based on Polar-HOGs Feature and Double Scales Direction Estimation

For resolving the problem of rotation-invariant human detection in natural scene, a rotation-invariant detection algorithm based on polar-HOGs and double-scale direction estimation is proposed in this paper. The algorithm first transforms rotation of object to cycle translation by using polar coordinate mapping, then eliminates the effect of rotation by applying reverse cycle translation to p-θ mapping according to candidate direction angle obtained by double-scale direction estimation, finaly employs SVM detector to finish the classification after extracting polar HOGs feature. The experimental results demonstrate that our algorithm is superior to the traditional algorithm when applied to human detection in the condition of non-rotation, and also be succeeding in multi rotation human classification and detection with high TPR against low FRP which is acceptable.