Prevalent human detection methods are usually applied in cases without rotation angle, and their detection rates are poor when rotation angle varies. In order to solve the issue, an algorithm which could identify human with variable rotation angle was proposed. Firstly, Radial Gradient Transform (RGT) method was adopted to obtain the rotation-invariance gradient. Then, adopting the method similar to the way that blocks were overlapped in the Histogram of Oriented Gradient(HOG) feature, a plurality of descriptors with rotation angle information were obtained and connected linearly into a descriptor group with rotation invariance feature, according to the descriptors' rotation angle. Finally, the human detection algorithm was conducted with the support of a two-level cascaded classifier based on Support Vector Machine (SVM). The recognition rate of the proposed algorithm achieves more than 86% for a human test set with 144 different rotation angles based on the INRIA pedestrian database. In the meantime, the false detection rate is less than 10% for a non-human test set with 144 different rotation angles. The experiments indicate that the proposed algorithm can be used for human detection in an image with arbitrary rotation angle.