This paper presents an experimental study on pedestrian classification and detection in far infrared (FIR) images. The study includes an in-depth evaluation of several combinations of features and classifiers, which include features previously used for daylight scenarios, as well as a new descriptor (HOPE -Histograms of Oriented Phase Energy), specifically targeted to infrared images, and a new adaptation of a latent variable SVM approach to FIR images. The presented results are validated on a new classification and detection dataset of FIR images collected in outdoor environments from a moving vehicle. The classification space contains 16152 pedestrians and 65440 background samples evenly selected from several sequences acquired at different temperatures and different illumination conditions. The detection dataset consist on 15224 images with ground truth information. The authors are making this dataset public for benchmarking new detectors in the area of intelligent vehicles and field robotics applications.