We investigate the detection of person in the omnidirectional images, adopting a linear SVM. We have implemented HOG-based descriptors, for omnidirectional and spherical images. In this paper we studied the influence of each parameter in our algorithm on the performances of person detections in catadioptric images. However, few studies have elaborated the problem of human detection using this type of cameras; therefore we have set up our own test base. Our results show that our detector can robustly detect people in omnidirectional images, as soon as the algorithm is adapted to the distortions introduced by the use of the omnidirectional camera.