基于深度学习的目标检测系统的研发

发布时间：2018-05-20 18:13

本文选题：目标检测 + 深度学习　；参考：《首都经济贸易大学》2017年硕士论文

【摘要】：计算机科学的飞速发展,给人类的生活带来了很大的进步,使人类的生活变得越来越智能。人工智能一直是人类孜孜不倦探索得重要领域。众所周知,人类的视觉是感知外部世界的重要组成部分,科学研究表明,人的百分之七八十的信息是通过视觉来感知到的。所以,在人类探索人工智能的漫漫长途中,计算机视觉一直都是一个重要的研究方向。计算机视觉涉及到图像处理,机器学习,模式识别等多个学科,最终目的是为了模拟人的视觉,以便用计算机完成各种识别任务。其中,目标检测是计算机视觉方向中非常重要的一个子方向。目标检测主要是检测出图片中所关注的目标,例如,自动驾驶系统对于目标检测的要求就是要检测出当前行车环境中的行人、车辆等各种物体。由于真实路况的复杂性,要求检测系统对于场景有着较高级别的语义理解。过去,大部分目标检测算法基本是基于传统的滤波方法,提取人工设计出来的经典特征,如SIFT[22],HOG[2],然后放入经典的分类器(如SVM[30]、Adaboost[29])进行分类识别。由于使用的是手工特征,所以鲁棒性较差,而且工作量大,当环境出现明显变化时,目标检测效果的差异很大。由于深度学习中卷积神经网络的极强的特征表达能力,提取的特征具有非常强的鲁棒性,所以,本文主要是利用了基于深度学习的比较经典的检测框架—Faster R-CNN[5],并在此基础上尝试着使用不同的特征提取层,在传统经典模型的基础上,对网络结构进行了改变,使现在的网络模型在精度和速度之间作了更好的权衡。并利用标定的数据对模型进行训练,调节参数,最终训练出一个精度和速度较好的模型,并应用到检测系统中。本文的目标检测系统的开发环境为Linux,利用专注图像界面的Qt图形界面库作为界面的开发框架,底层使用了C++语言。本文中所描述的目标检测系统开发过程主要包括整体的需求分析、总体的设计与实现和测试等。最后通过测试,证明系统在硬件和性能上都有着良好的表现。
[Abstract]:The rapid development of computer science brings great progress to human life and makes human life more intelligent. Artificial intelligence has always been an important field for human beings to explore tirelessly. As we all know, human vision is an important part of the perception of the external world. Scientific research shows that 70% of human information is perceived through vision. Therefore, computer vision has always been an important research direction in the long-distance exploration of artificial intelligence. Computer vision involves many subjects, such as image processing, machine learning, pattern recognition and so on. Among them, target detection is a very important sub-direction in the direction of computer vision. Target detection is mainly to detect the object concerned in the picture. For example, the requirement of automatic driving system for target detection is to detect all kinds of objects such as pedestrians, vehicles and so on in the current driving environment. Because of the complexity of the real road conditions, the detection system is required to have a higher level of semantic understanding of the scene. In the past, most of the target detection algorithms were based on traditional filtering methods to extract the classical features, such as SIFT [22] Hog [2], and then put them into classical classifiers (such as SVM [30] / Adaboost [29]) for classification and recognition. Because the manual feature is used, the robustness is poor, and the workload is large. When the environment changes obviously, the target detection effect is very different. Because of the strong feature expression ability of convolution neural network in deep learning, the extracted feature is very robust. In this paper, we mainly use the more classical detection framework based on depth learning-Faster R-CNN [5], and on this basis try to use different feature extraction layers, on the basis of the traditional classical model, the network structure is changed. So that the current network model to make a better balance between accuracy and speed. The calibration data is used to train the model and adjust the parameters. Finally, a model with good precision and speed is trained and applied to the detection system. The development environment of the target detection system in this paper is Linux. the QT graphical interface library which focuses on the image interface is used as the development framework of the interface, and C language is used in the bottom layer. The development process of the target detection system described in this paper mainly includes the whole requirement analysis, the overall design and implementation, and the test and so on. Finally, through the test, it is proved that the system has good performance in both hardware and performance.
【学位授予单位】：首都经济贸易大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP18

【参考文献】