| Budget Amount *help |
¥2,400,000 (Direct Cost: ¥2,400,000)
Fiscal Year 2005: ¥1,300,000 (Direct Cost: ¥1,300,000)
Fiscal Year 2004: ¥1,100,000 (Direct Cost: ¥1,100,000)
|
|---|
| Research Abstract |
An effective navigation system is an ultimate necessity to improve the mobility of 37 millions of blind people all over the world. The objective of this study is to develop an autonomous travel aid for the blind based on computer vision. For the purpose of achieving such an objective, we developed a device with which the blind would be able to autonomously detect important information for safely negotiating a pedestrian crossing. The important information is as follows : 1) location of a pedestrian crossing, 2) length of a pedestrian crossing, 3) traffic light color, and 4) existence of a vehicle. Using image data observed with a single camera, we proposed methods for image analysis of a pedestrian crossing scene 1) to detect the location of a pedestrian crossing, 2) to measure the length of a pedestrian crossing, 3) to detect a traffic light color, and 4) to detect the existence of a vehicle in the following.
1) Detection of location of crossing : Two noble approaches were developed fo
… More
r detection of the location of crossing. One approach was based on projective invariants using image feature points extracted by Fisher criterion. The second approach was based on bipolarity feature. Both methods gave more than 95% accuracy with no false positive.
2) Measurement of crossing length : Two new methods were developed for the measurement of crossing length. One was vector geometry method and the other was band parameter method. In a pedestrian crossing, the usual black road surface is painted with constant width periodic white bands. In vector geometry method, the crossing length was estimated using the left- and the right-border lines, the first-, the second- and the end-edge lines of the crossing region. In band parameter method, the crossing length was estimated from the number of crossing bands within the crossing region along with band width parameter. The rms error was 2m for both the methods.
3) Detection of traffic light color : The color distribution of similarity for a true traffic light color was prepared and stored as a reference datum. From the reference of similarity distribution, the candidate regions for traffic light in an image were extracted. From the candidate regions, a true traffic light region was selected by using an Affine moment invariant. The proposed method was successful in detecting traffic lights with an accuracy of 94%.
4) Detection of a vehicle : Candidate regions for vehicles were extracted using Gabor filter outputs. From the candidate regions, a true region with a vehicle was selected. A vehicle was detected by checking the selected region with an Affine moment invariant. Less
|
