Vision-based car counting for multi-story carparks
Method and apparatus for counting vehicles at entrances, exits and transition zones of multi-story carparks, including particularly where the ceiling heights can be just marginally higher than the tallest expected vehicle, with a view to determine the carpark occupancy at different carpark levels by counting passing vehicles using a vision-based car counting system without relying on viewing test patterns or employing a blocking beam scheme and yet tolerating vehicles transgressing partially or fully into the wrong lane of a two-lane two-way road while ignoring vehicles moving opposite to the expected direction. Without imposing additional constraints to ambient carpark illumination, the methodology copes with highly specular vehicle surfaces, ignores non-vehicular objects and detects moving cast shadow or highlight, and adapts to daily and seasonal scene changes, and yet estimates vehicle speed.
1. A method for counting multiple passing vehicles in transition zones or entrances or exits of a multi-story carpark in a given direction along a two-lane two-way traffic road beneath a low clearance ceiling of the carpark, the method comprising:
providing a video camera for producing a video stream, the video camera including a wide angle rectilinear lens forming a pyramidal view volume of the video camera, the pyramidal view volume having a field of view and four outer faces;
locating the video camera at the ceiling above a road side edge so as to image the road across the field of view and align one of the faces of the pyramidal view volume parallel to the road side edge and perpendicular to the road;
processing the video stream to model a prevailing background model of the road in the field of view;
defining a rectangular monitor zone extending across the road;
deriving a smoothed incident image for the rectangular monitor zone; and
performing directional gradient processing on the smoothed incident image to produce a foreground image, comprising the steps of:
convolving the smoothed incident image with Sobel horizontal and vertical kernels to produce horizontal and vertical gradient images, respectively;
computing a gradient magnitude image from the horizontal and vertical gradient images;
computing an adaptive suppression threshold for suppressing noise in the gradient magnitude image;
comparing the smoothed incident image to the prevailing background model on a pixel-by-pixel basis; and
producing the foreground image in response to determining a gradient orientation conformance between the smoothed incident image and the prevailing background model and in response to determining an intensity differential between the smoothed incident image and the prevailing background model;
deriving a motion direction from a temporal change in the foreground image; and
altering a vehicle count in response to deriving a motion direction.
2. The method of claim 1 , wherein the step of deriving a smoothed incident image for the rectangular monitor zone includes low pass filtering the video stream with a Gaussian point spread function.
3. The method of claim 1 , wherein:
the smoothed incident image consists of pixels; and
the step of deriving a smoothed incident image for the rectangular monitor zone comprises low pass filtering each of the pixels of the smoothed incident image with a Gaussian point spread function.
4. The method of claim 3 , wherein the step of performing directional gradient processing on the smoothed incident image to produce a foreground image includes performing directional gradient processing on each of the pixels of the smoothed incident image.
5. A method for counting multiple passing vehicles in transition zones or entrances or exits of a multi-story carpark in a given direction along a two-lane two-way traffic road beneath a low-clearance ceiling of the carpark, the method comprising:
providing a video camera for producing a video stream, the video camera including a wide angle rectilinear lens forming a pyramidal view volume of the video camera, the pyramidal view volume having a field of view and four outer faces;
locating the video camera at the ceiling above a road side edge so as to image the road across the field of view and align one of the faces of the pyramidal view volume parallel to the road side edge and perpendicular to the road;
processing the video stream to model a prevailing background model of the road in the field of view;
processing the video stream to update the prevailing background model of the road so as to reveal a foreground image, comprising the steps of:
deriving a smoothed incident image;
comparing the smoothed incident image to the prevailing background model on a pixel-by-pixel basis; and
updating a pixel of the prevailing background model with a pixel of the smoothed incident image to reveal the foreground image, in response to the pixel of the prevailing background model differing from the pixel of the smoothed incident image by a threshold;
wherein the threshold is dependent on a difference between a gradient orientation of the pixel of smoothed incident image and a gradient orientation of the pixel of the prevailing background model;
deriving a motion direction from a temporal change in the foreground image; and
altering a vehicle count in response to deriving a motion direction.
6. The method of claim 5 , wherein the step of processing the video stream to update the prevailing background model of the road so as to reveal a foreground image further comprises the steps of:
defining a rectangular monitor zone extending across the road; and
deriving the smoothed incident image at the rectangular monitor zone.
7. The method of claim 6 , wherein the step of deriving the smoothed incident image at the rectangular monitor zone includes low pass filtering the video stream with a Gaussian point spread function.
8. The method of claim 6 , wherein:
the smoothed incident image consists of pixels; and
the step of deriving the smoothed incident image at the rectangular monitor zone comprises low pass filtering each of the pixels of the smoothed incident image with a Gaussian point spread function.
9. The method of claim 5 , further comprising relaxing the threshold in response to conformity between the gradient orientations of the pixels of the smoothed incident image and the prevailing background model.
10. The method of claim 5 , further comprising making the threshold stringent in response to non-conformity between the gradient orientations of the pixels of the smoothed incident image and the prevailing background model.