The right figure shows the sequence of events in our proposed method. A video is made of the area in front of the vehicle by means of the in-vehicle camera. Symmetry is judged from the left and right images of the pedestrian by means of high-speed processing, and the pedestrian candidate area is narrowed down. The edge detector is processed, and binarization is carried out on the source image. In this way, it is possible to scan the extraction window in the image. The current method for capturing a pedestrian's shape employs co-occurrence histograms of oriented gradients (CoHOG). In this research, however, we propose the use of Extended Co-occurrence Histogram of Oriented Gradient (ECoHOG) as an improvement of CoHOG. ECoHOG holds particular advantages over previous methods. The first method is edge-magnitude accumulation. The second method is normalisation.

  We verified the pedestrian-detection accuracy using monocular camera. We used precision, recall and the F measure as indicators of detection accuracy, whereby precision indicates the accuracy of the system, recall is the detection rate of system in an actual pedestrian area, and the F measure is calculated as the harmonic mean from precision and recall. In experiment, the newly proposed method -ECoHOG- gave a higher value (0.6795) in the F measure than CoHOG (0.6321). This indicates that ECoHOG is superior in terms of accuracy. We believe this accuracy derives from the edge-magnitude accumulation and normalization of the ECoHOG method. ECoHOG can correctly identify pedestrians and the background in cases where the background includes object that have a similar edge-magnitude accumulation.