A method for detection of a target present on at least two images of the same scene captured by separate cameras comprises a prior step for learning about targets under setup conditions and further comprises, under conditions of use, a simultaneous classification step for objects present on the images, the target being said to be detected as soon as an object is classified as being one of the targets learned during the learning step. The classification step includes a step for adapting at least one of the images to the setup conditions under which the learning step took place. Application: surveillance, assistance and safety based on stereoscopic images.