Vision‐based pedestrian detection for intelligent vehicles applications is a crucial and active research area due to the essential benefits in terms of reducing the number of accidents involving pedestrians and vehicles. During the last decade a considerable amount of research studies have been proposed, filling the gap between prototypes and commercial implementations. Pedestrian detection systems can be roughly divided into three main different sub‐parts: Region Of Interest – ROI – selection, classification and tracking. Previous surveys have covered the literature in a holistic way. An example would be, analyzing all the solutions proposed for all the stages and including higher level analysis, but in most cases they give more emphasis to the classification stage. Due to the difficulty of this detection task, the variety of solutions, sensor configurations (monocular/stereo; visible/infrared) available in the literature, we propose to break down the variability of the problem by providing exhaustive review of one specific stage: stereo‐based ROI selection. ROI selection is a key component that has to be designed to provide generic obstacles at lowest false negative rate and maintain a low number of false positives. The number of missed pedestrians has to be approximately equal to 0 since a pedestrian missed by the ROI selection stage would not be detected in further stages. In addition, the number of non‐pedestrians obstacles should be as low as possible to reduce both the number of false alarms and the computational costs of further stages. In contrast to monocular approaches, stereo ROI selection determines the relative distance between the pedestrian and the vehicle, assuring that the reported candidates are related with real physical objects. The stereo‐based ROI selection step is also divided into different components that are independently analyzed, increasing visibility for future proposals and developments. Discussion is finally presented highlighting the current problems for obtaining a global overview of the actual performance of the different approaches and analyzing future trends.
