Please use this identifier to cite or link to this item:
|Title:||Video surveillance system for human detection||Authors:||Xu, Wanxin||Keywords:||DRNTU::Engineering::Electrical and electronic engineering||Issue Date:||2017||Abstract:||The problem I need to solve is the pedestrian detection in campus monitor. Computer vision studies based on deep learning algorithm provide relatively precise result for frame and video detection. Among all kinks of deep learning frameworks and algorithms, Caffe and Faster R-CNN perform outstandingly in both detection speed and accuracy rate. In this dissertation, Caffe and Faster R-CNN are applied on both CPU and GPU to detect the people in campus station video based on VGG network and ZF network. To visualized display the detection process and result, I created an interface using python. In the interface, functions of video selection, algorithm detection, result displaying are integrated. After the test of algorithm, the mean average precision for people detection is around 0.76. In the detection of campus monitor, most of complete pedestrians with proper size can be detected. The original request of the dissertation is satisfied. To further improve the performance of the algorithm, the network is trained using VOC 2007, VOC2012, VOC 2007+VOC 2012 and the superposition of other labeled pedestrian datasets. After the enhance training, the number of detected objects in campus monitor video increased. These results proved that increasing training samples of a specific class contributes to the performance of Faster R-CNN algorithm.||URI:||http://hdl.handle.net/10356/72557||Fulltext Permission:||restricted||Fulltext Availability:||With Fulltext|
|Appears in Collections:||EEE Theses|
Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.