Neural network-based vehicle detection system and method

• US 5,448,484 A
• Filed: 11/03/1992
• Issued: 09/05/1995
• Est. Priority Date: 11/03/1992
• Status: Expired due to Fees

• Current Assignee: Chris T. Hendrickson, Darcy M. Bullock, James H. Garrett Jr
• Original Assignee: Chris T. Hendrickson, Darcy M. Bullock, James H. Garrett Jr
• Inventors: Bullock, Darcy M., Garrett, James H. Jr., Hendrickson, Chris T.
• Primary Examiner(s): Black, Thomas G.
• Assistant Examiner(s): Wieland, Susan
• Application Number:
  US07/970,800
• Time in Patent Office: 1,036 Days
• Field of Search: 364/436, 395/23, 340/933, 340/934, 340/937
• US Class Current: 701/117
• CPC Class Codes: G06F 18/24143 Distances to neighbourhood prototypes, e.g. restricted Coulomb energy networks [RCEN]
• CPC Class Codes: G06T 2207/10016 Video; Image sequence
• CPC Class Codes: G06T 2207/20084 Artificial neural networks [ANN]
• CPC Class Codes: G06T 2207/20104 Interactive definition of region of interest [ROI]
• CPC Class Codes: G06T 2207/30236 Traffic on road, railway or crossing
• CPC Class Codes: G06T 7/70 Determining position or orientation of objects or cameras camera calibration G06T7/80
• CPC Class Codes: G06V 10/454 Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
• CPC Class Codes: G06V 10/82 using neural networks
• CPC Class Codes: G06V 20/54 of traffic, e.g. cars on the road, trains or boats
• CPC Class Codes: G08G 1/04 using optical or ultrasonic detectors

Abstract

The present invention is directed to a neural network-based system for detecting the presence of a vehicle within a traffic scene. The vehicle detection system comprises an apparatus for producing an image signal representative of an image of the traffic scene and a trainable neural network for identifying the presence of a vehicle within the traffic scene. The present invention is also directed to a method for detecting the presence of a vehicle within a traffic scene. The vehicle detection method includes the steps of producing an image signal representative of an image of the traffic scene, collecting a training set of these image signals, training a neural network from this training set of image signals to correctly identify the presence of a vehicle within the traffic scene and performing surveillance of the traffic scene with the trained neural network to detect the presence of a vehicle.

Claims

1. A system for detecting the presence of a vehicle within a traffic scene, said vehicle detection system comprising:

means for producing an image signal representative of an image of the traffic scene that includes a video camera for producing an analog video signal representative of a video image of the traffic scene; and

means for processing said image signal to identify the presence of the vehicle within the traffic scene wherein said means for processing includes;

means for producing a digitized image signal from said analog video signal;

means for specifying a detection zone within said video image and means for extracting a portion of said digitized image signal corresponding to said detection zone to produce a digitized detection zone signal;

means for producing a tiled detection zone pixel map from said digitized detection zone signal;

a trainable neural network comprising input processing units, hidden processing units and output processing units wherein an output from each of said input processing units is connected to an input of each of said hidden processing units and an output from each of said hidden processing units is connected to an input of each of said output processing units and wherein said trainable neural network produces an output signal at each of said output processing units representative of the presence of the vehicle within a portion of said detection zone and wherein an ordered vector of said output signals produces a Gaussian-shaped curve;

means for inputting said tiled detection zone pixel map into said trainable neural network; and

an output filter for producing a presence output signal indicating that the vehicle is within said detection zone, said output filter including means for producing a pulse output signal indicating that the vehicle has passed through said detection zone by performing a matched filter operation on a time series of said ordered vectors of said output signals to detect when the vehicle has passed through said detection zone.

2. The vehicle detection system of claim 1 wherein said trainable neural network is comprised of a single layer of input processing units, more than one layer of hidden processing units and a single layer of output processing units and wherein an output from each of said input processing units is connected to an input of each of said hidden processing units in a first layer of said hidden processing units, an output from each of said hidden processing units in a last layer of said hidden processing units is connected to an input of each of said output processing units, and multiple layers of said hidden processing units are interconnected such that said output from each of said hidden processing units in any one but said last of said layers of said hidden processing units is connected to said input of each of said hidden processing units in a next layer of said hidden processing units.

3. The vehicle detection system of claim 1 wherein said output filter includes means for producing a presence output signal when one of said output signals exceeds a threshold value.

4. The vehicle detection system of claim 3 wherein said threshold value is in the range from 0.6 to 0.8.

5. The vehicle detection system of claim 1 wherein said trainable neural network is implemented in computer software as a neural network simulator running on a computer.

6. The vehicle detection system of claim 1 wherein said trainable neural network is implemented in computer hardware.

7. The vehicle detection system of claim 1 wherein a numerical connection weight is assigned to (i) each of said connections between each of said outputs of each of said input processing units and each of said inputs of each of said hidden processing units and (ii) each of said connections between each of said outputs of each of said hidden processing units and each of said inputs of each of said output processing units.

8. The vehicle detection system of claim 7 wherein a numerical bias is assigned to each of said hidden processing units and each of said output processing units.

9. The vehicle detection system of claim 8 wherein the value of each of said numerical connection weights and each of said numerical biases is determined through a closed-loop training procedure utilizing backpropagation techniques.

10. The vehicle detection system of claim 9 wherein said closed-loop training procedure is the Generalized Delta Rule.

11. A method for detecting the presence of a vehicle within a traffic scene, comprising the steps of:

producing an image signal representative of an image of the traffic scene and an analog video signal representative of a video image of the traffic scene;

processing said image signal to identify the presence of the vehicle within the traffic scene;

producing a digitized image signal from said analog video signal;

specifying a detection zone within said video image and extracting a portion of said digitized image signal corresponding to said detection zone to produce a digitized detection zone signal;

producing a tiled detection zone pixel map from said digitized detection zone signal;

inputting said tiled detection zone pixel map into a trainable neural network comprising input processing units, hidden processing units and output processing units wherein an output from each of said input processing units is connected to an input of each of said hidden processing units and an output from each of said hidden processing units is connected to an input of each of said output processing units and wherein said trainable neural network produces an output signal at each of said output processing units representative of the presence of the vehicle within a portion of said detection zone and wherein an ordered vector of said output signals produces a Gaussian-shaped curve;

producing a presence output signal indicating that the vehicle is within said detection zone; and

producing a pulse output signal indicating that the vehicle has passed through said detection zone by performing a matched filter operation on a time series of said ordered vectors of said output signals to detect when the vehicle has passed through said detection zone.