Pedestrian detector
A pedestrian detector includes an image-taking mechanism and an image analyzer. The image-taking mechanism is provided on a vehicle and used to take an image in front of the vehicle. The image analyzer is connected to the image-taking mechanism and used to receive the image from the image-taking mechanism and analyze the image to determine whether the image conforms to a pedestrian to determine whether there is a pedestrian in front of the vehicle.
1. A pedestrian detector including
a camera provided on a vehicle and arranged to take an image in front of the vehicle; and
an image analyzer connected to the camera so as to receive the image from the camera and comprising:
a composite feature-extracting module comprising:
a histogram of oriented gradients (HOG) feature-extracting module that extracts contour features of the image;
a granule-based comparison feature-extracting module that extracts variations of brightness in an array of granules of the image with granules considered both individually and as combined with other granules and creates a 10-dimensional vector to overcome errors related to poor quality of the image; and
a feature-combination module that combines the extracted contour features of the image with the extracted variations of brightness in granules of the image and create a single feature vector that includes the extracted contour features and the extracted variations of brightness in granules of the image; and
a support vector machine that determines whether there is a pedestrian based on the single feature vector and whether there is a pedestrian in front of the vehicle.
2. The pedestrian detector according to claim 1 , wherein the granule-based comparison feature-extracting module includes
a granule brightness-averaging unit that averages the brightness of each granule; and
a-gradient-histogram-encoding unit that produces gradient features of the granules.
3. The pedestrian detector according of claim 1 , wherein a size of dimensions of the single feature vector is equal to a sum of a size of dimensions of HOG and a size of dimensions of the granule-based gradient features.
4. The pedestrian detector of claim 1 , wherein the array of granules is a 4 by 4 square array and the granules are processed by the granule-based comparison feature-extracting module as a plurality of 2 by 2 arrays, each 2 by 2 array translated by one granule from a previous 2 by 2 array and each 2 by 2 array resulting in a one dimensional vector such that a 9 dimensional vector is produced for the 4 by 4 array.
5. The pedestrian detector of claim 4 , wherein four non-overlapping 2 by 2 arrays are each treated as four larger granules and the four larger granules are processed to obtain an additional one dimensional vector so as to produce the 10 dimensional vector.