Pedestrian detection using fast region-based convolutional neural network method

• Main Author: Wang, Xiaoxu
• Other Authors: Lu Yilong
• Format: Final Year Project
• Language: English
• Published: 2019
• Subjects: DRNTU::Engineering::Electrical and electronic engineering
• Online Access: http://hdl.handle.net/10356/77882
• Institution: Nanyang Technological University

People are the center of all kinds of social activities. In real-life scenario, people are the most important objects of concern, such as pedestrians crossing the road, security inspections, etc. As a specific object detection method, pedestrian detection is the premise of vehicle-assisted driving, intelligent video surveillance, as well as human behavior analysis. With the continuous enhancement of hardware capabilities and related proposed algorithms, the performance of the pedestrian detection system is also constantly improving. Pedestrian detection possesses very important research significance and value, thus it has attracted wide attention of researchers as compared to the past. Singapore is one of the leading countries in Asia that invested numerous recourses in developing and researching on Artificial Intelligence (AI), such as pedestrian detection on self-governing vehicles. Many educational institution and research centers including NUS, NTU have developed their own prototypes of autonomous vehicle which have tested within the school campus. Indeed, Pedestrian detection will gradually become an indispensable research topic in Singapore, and in the near future, the research achievement will be applied to people's daily lives. In this thesis, the author will touch on the evolution of the new and old methods of object detection. By reviewing some literature researches with analysis, it is found that the existing Faster RCNN method has replaced the Fast RCNN method which was proposed earlier for this project. Furthermore, Faster RCNN method will be able to perform better detection function and deliver more accurate detection results for pedestrian detection. As a result, the construction and working principle of Faster R- CNN method which used in the experiment, and the corresponding dataset will be introduced in detail. Lastly, a demo pedestrian detection program will be developed based on Faster R-CNN platform with VGG16 network trained for detection on PASCAL VOC 2007. Comparisons and analysis on experiment results using different training models with variable training iterations will be discussed in this paper.