基于神经网络的行人检测方法研究

发布时间：2018-02-25 20:27

本文关键词： 行人检测图像处理神经网络回归窗口全局概率　出处：《西北大学》2017年硕士论文　论文类型：学位论文

【摘要】：行人检测是计算机视觉中一个很有挑战性的领域。其应用广泛如自动驾驶、智能监控、智能机器人等。近些年,基于神经网络的行人检测方法取得了巨大的发展,大量的神经网络模型被提出并用于行人检测领域。但是,现有基于神经网络方法的检测速度满足不了实时性要求。因此为了提高检测速度,本文对几种经典的神经网络行人检测方法进行了深入分析,研究了回归窗口和全局概率的行人检测方法。本文主要工作如下:分别从提高准确度与提高速度两方面对现有行人检测方法进行了对比分析,发现鲜有方法能在保证较高准确度的同时兼顾检测速度。鉴于回归参数检测方法计算高效,本文提出基于回归窗口与全局概率的行人检测方法。首先包含两个网络模型:回归窗口网络与全局概率网络,回归窗口网络负责生成固定个数的窗口以及计算窗口局部概率,全局概率网络负责估计行人窗口的空间概率。其次,为了级联训练两个网络模型的参数,基于联合优化的思想,本文进一步探索端到端的网络结构与传播方式,以保证较高的检测速度;同时,研究全局概率增强回归窗口准确度的方法,以保证较高的检测准确度。最后,在Caltech公共数据集上进行实验,实验结果表明,本文方法与相同准确度的方法相比,检测速度高十倍左右;与相同速度的方法相比,检测准确度提高两倍左右。
[Abstract]:Pedestrian detection is a challenging field in computer vision. It is widely used in automatic driving, intelligent monitoring, intelligent robot and so on. In recent years, neural network-based pedestrian detection methods have made great progress. A large number of neural network models have been proposed and used in pedestrian detection field. However, the current detection speed based on neural network method can not meet the real-time requirements, so in order to improve the detection speed, In this paper, several classical neural network pedestrian detection methods are deeply analyzed. The methods of pedestrian detection based on regression window and global probability are studied. The main work of this paper is as follows: comparing and analyzing the existing pedestrian detection methods from the aspects of improving accuracy and improving speed respectively. It is found that there are few methods that can guarantee high accuracy and at the same time take into account the detection speed. In this paper, a pedestrian detection method based on regression window and global probability is proposed. Firstly, there are two network models: regression window network and global probability network. The regression window network is responsible for generating a fixed number of windows and calculating the local probability of window. The global probabilistic network is responsible for estimating the spatial probability of pedestrian window. Secondly, in order to concatenate the parameters of the two network models, based on the idea of joint optimization, this paper further explores the end-to-end network structure and propagation mode. At the same time, the method of global probability to enhance the accuracy of regression window is studied to ensure the higher detection accuracy. Finally, the experiment is carried out on the common data set of Caltech, and the experimental results show that, Compared with the method of the same accuracy, the detection speed of this method is about ten times higher than that of the method of the same speed, and the accuracy of detection is about two times higher than that of the method of the same speed.
【学位授予单位】：西北大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【相似文献】