FPGA-based pedestrian detection under strong distortions

Pedestrian detection is one of the most popular computer vision challenges in the automotive, security and domotics industries, with several new approaches and benchmarks proposed every year. All of them typically consider the pedestrians in a standing pose, but this assumption is not always applicable. It is the case of embedded camera systems used for crowd monitoring or in driving assistance systems for big vehicles maneuvering. Such systems are commonly installed as higher as possible and make use of fish-eye lenses to provide a top and wide field of view. Actually, such configurations introduce both perspective and optical distortions in the image that, even when corrected, still provide stretched silhouettes that can hardly be detected by cutting-edge pedestrian detection algorithms. In this paper we focus on this scenario, showing (a) that one of the most effective models for pedestrian detection, that is the Deformable Part Model (DPM), can be efficiently implemented in FPGA to dramatically speed up the computation, and (b) how it can be modified for dealing with highly distorted pictures of humans. The resulting framework, dubbed Deformable Part Model for Local Spatial Deformations (DPM-LSD), gives convincing figure of merits in terms of accuracy and throughput, on a new top-view fish-eye based pedestrian dataset (dubbed Fish-Eyed Pedestrians), also comparing with widely-used competitors (standard DPM and Dalal-Triggs).