Fast Pedestrian Detection and Tracking System based on Enhanced Histogram of Gradient Features and Efficient Object Classification Techniques

• Main Authors: Ming Chen, 成明
• Other Authors: Yen-Lin Chen
• Format: Others
• Language: zh-TW
• Online Access: http://ndltd.ncl.edu.tw/handle/vz9m7f

碩士 === 國立臺北科技大學 === 資訊工程系研究所 === 104 === In the field of object detection, most researchers have employed Histogram of Oriented Gradients (HOG) as the main feature for objects with complex edge. Although HOG is very good at describing edges, it has a fatal flaw, that is, HOG demands a lot of computing power so that real-time detection capability is highly challenged. Accordingly, many real-time applications do not make use of HOG as describing main feature. In this thesis, we propose an enhanced HOG which aims to achieve real-time performance. The proposed approach can reduce a lot of calculation steps in calculating HOG features which thus outperforms conventional methods with only 0.001 error degree. With this enhancement, the proposed method can calculate HOG features for whole frame in a very short time, so that it can achieve real-time detection. Moreover, this thesis adopts a faster SVM classifier as the core classifier for classifying all the input data. The conventional SVM classifier difficultly achieve real-time classification, due to enormous computational cost. The other contribution of the thesis is that our proposed method is about 43-times faster than the combination of original HOG and SVM classifier. Finally, to make our approach more applicable in various scenarios, our algorithm is implemented onto the embedded system NVIDIA TX1. The experimental results depict that the detection rate can reach 95% with the accuracy rate 99%.