Efficient DPM-based Pedestrian Detection Using Shift with Importance Sampling

• Main Authors: Hsiao, Chia-Jen, 蕭家任
• Other Authors: Hsieh, Jun-Wei
• Format: Others
• Language: zh-TW
• Published: 2017
• Online Access: http://ndltd.ncl.edu.tw/handle/x6affa

碩士 === 國立臺灣海洋大學 === 資訊工程學系 === 105 === Pedestrian detection is an importantly safe work that can be used in self-driving assistance systems to help drivers ensure the movement of pedestrian in front. It can also be used in many dangerous workplaces to prevent pedestrians or staffs from operating the equipment to cause danger to others. This technology can be used in many aspects of intelligent monitoring, such as at junctions, public places, safe maintenance and commercial advertising. To achieve the above objectives, we use the pedestrian detection system based on the Deformable Parts Model algorithm as an architecture. In order to improve the efficiency of pedestrian detection, we use Shift with Importance Sampling scanning scheme to quickly find the correct location of each pedestrian with minimum tries and tests. During the two years of the study, we found that the use of depth learning to do the classification effect is very remarkable but requires the support of the equipment to achieve immediate detection. In this paper, we propose the technology can be applied to the general PC equipment. In order to ensure that our pedestrian detection system can maintain a certain accuracy and immediacy, we have tried a lot of methods and then found a combination of high accuracy and immediacy way.