| Summary: | 碩士 === 國立成功大學 === 民航研究所 === 104 === As the rapid development of transportation nowadays, distracting from roads causing pedestrian bumping accidents happened from times to times. New ways of reducing driver’s workload and avoiding collisions are popping up fast. Furthermore, unmanned driving system evolves revolutionary. All of the factors above show that the detection of pedestrians is much more important than ever. This thesis develops a stereo vision system to detect and localize pedestrians. First, a two-camera system is calibrated in order to make sure the alignment of the images took from both cameras. Second, frame segmentation is applied. By segmentation, areas with no pedestrian can be ignored to reduce computational time. Then multi-scale Histogram of Oriented Gradient (HOG) descriptors are extracted to detect pedestrians using Support Vector Machine (SVM) as the classifier. The detected pedestrians are boxed to label their location in the image. The Oriented FAST and Rotated BRIEF (ORB) feature algorithm is then used to estimate the disparity information, which can be used to determine the relative distance between pedestrians and the vehicle. Finally, applying weightings to reduce the errors made by wrong matched points. Several examples are provided to demonstrate that the proposed system could be used to detect pedestrians and have possible applications in traffic collision avoidance systems.
|
|---|
