Acceleration of HOG based pedestrian detection in FIR camera video stream

Pedestrian detection is one the main problem for the automotive applications which is a challenging task to do. In recent years, the number of approaches was developed to speed up the detection rate and accuracy. However, the general problem of detectors remains. The aim of the paper was to evaluate experimentally current and efficient pedestrian detection methods in night vision applications and propose some modifications to accelerate the image analysis workflow. In this paper, for night vision application we used an FIR domain camera. The novelty of the proposed solution lays in the application of HOG based pedestrian detector for FIR domain camera. An acceleration of the algorithm was achieved using subtraction of the thermally active regions before supplying these regions to the pre-trained feature descriptor. An experimental investigation has shown the significant improvement in pedestrian detection speed using solution, proposed in this paper. Our study shows that state of the art detectors can gain nearly triple initial detection rate using the same image data.