We present recent experiments on BEST-2 images concerning detection and classification of vehicles on Infra-Red images. The approach is based on learning by example and it involves Neural Network techniques. Concerning detection, the image is analysed by multi-resolution scanning. For each resolution and each window location, the content of the window is normalized in size, average brightness, and contrast. Then, it is propagated through a multi-layer neural network, which provides a decision ("vehicle" or "background") and a confidence measure. All the decisions are then fused across the various resolutions in order to suppress multiple detections at the same location. The fusion is based on analysis of confidence. Concerning classification, 6 classes have been defined: helicopter, truck, and 4 classes of tanks. A neural network has been trained to recognize these classes independently of vehicle orientation and environment conditions. The algorithms descriptions is presented and experimental results are provided.