基於CENTRIST特徵和隨機過程實現行人偵測

Abstract

A method of pedestrian detection based on CENTRIST descriptor and stochastic process is proposed in this thesis. In related work such as C4 and Peng’s method, they use only single image as input, regardless driving is a continuous process. In our work, we will use sequential data and use stochastic process to help determine the possibility of pedestrian appearance. We use the training set cut from our own database built by driving recorder Papago P3 to train SVM models to be our basic object detector. Our experimental results show that our method outperforms C4 and Peng’s method in execution time and comparable accuracy by applying stochastic determination.