Research On Pedestrian Detection Method Based On Migration Learning

Pedestrian detection is the core of the application of intelligent monitoring systems, intelligent transportation systems and intelligent vehicle systems. The goal of the study of pedestrian detection methods based on transfer learning, is to improve the detection performance of the classifier by making use of transfer learning.This paper seperates the study of pedestrian detection based on transfer learning into two parts:the first one is sample selection strategy, and the second one is sample filtering algorithm. This paper also proposed a pedestrian detection model based on transfer learning. The main research of sample selection strategy, is to select samples for the training sample sets. Its goal is to enhance the detection performance of classifier. Sample filtering algorithm is mainly in the field of research methods to obtain the target sample from the detection results, so that to continue acquiring right new sample in the target field.This paper studies the sample selection strategy based on determined positive and determined negative samples. The main research study on two sample selection strategies:the first one is adding samples directly and the last is adding samples relies on a filtered set based on the original classifier. With the research, a conclusion is draw that the strategy which adding samples relies on a filtered set based on the original classifier is a relatively viable strategy at Caltech pedestrian database with a proper sample-grouping condition.This paper also studied sample filtering algorithm for transfer learning based on perdestrian size model. According camera imaging characteristics, this paper proposed pedestrian size model to meet the needs of specific scenarios captured by car camera. With Caltech pedestrian database, we verify the correctness of the proposed pedestrian size model.Finally, we studies the sample filtering algorithm on motion detection. With the imaging characteristics of car camera, we put forward the mode optical flow method. And on the Caltech pedestrian database specific consecutive frames were analyzed and it proves the mode optical flow method can be used for transfer learning to selecte samples on video by car camera under specific scenarios.