基于深度卷积神经网络的道路场景物体检测方法研究

发布时间：2018-02-28 09:50

本文关键词： 深度卷积神经网络 Faster R-CNN R-FCN SSD 物体检测　出处：《北方工业大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着经济的蓬勃发展,汽车已成为人们出行的主要交通工具。然而,在汽车拥有量迅速增长的同时,交通安全问题也日益突出。其中驾驶员注意力分散和疏忽是导致交通事故的主要因素。汽车车载辅助安全系统能够发现车辆周边潜在风险并向驾驶员预警的有效系统和工具,可有效提高驾驶的安全性,降低驾驶过程的风险。自动驾驶技术是集定位、感知、决策、控制等众多技术于一体的复杂系统。感知系统为自动驾驶或辅助驾驶系统提供了基础支撑数据,该系统的有效性成为整个自动驾驶系统有效性的基础保障。本课题基于深度卷积神经网络设计了一个针对道路场景条件下的物体检测器模型,用于检测对驾驶系统具有重要价值的目标。目前基于深度卷积神经网络的物体检测框架分为两个研究方向。一种是利用区域推荐方法与深度卷积神经网络特征提取方法相结合搭建物体检测框架;另一种是直接从深度卷积神经网络中回归出区域框和物体识别。本课题将对两种方法做进一步算法的研究,并讨论在不同方法下训练及测试道路场景物体检测器模型性能。本课题设计的道路场景物体检测器利用深度卷积神经网络的卷积层提取图像数据表面的特征,然后利用区域框提取算法定位道路场景物体在图像中的具体位置,再利用卷积层提取的特征识别物体的具体类别。最后将在高性能计算机下训练好的检测器模型植入嵌入式开发系统,并测试道路场景物体检测模型对于白天与夜间不同光照环境下道路物体检测并获得较好检测结果。对道路场景中的行人检测精度为83%,对于道路场景中小轿车、公共汽车、摩托车、自行车等各种车辆的检测平均精度为81%。
[Abstract]:With the rapid development of economy, cars have become the main means of transportation for people to travel. However, with the rapid growth of car ownership, Traffic safety problems are also becoming more and more prominent. Driver distraction and negligence are the main factors leading to traffic accidents. Vehicle aided safety system can detect the potential risk around the vehicle and warn the driver of the effective system and tools. It can effectively improve the safety of driving and reduce the risk of driving process. Autopilot technology is a set of positioning, perception, decision-making, A complex system in which many technologies such as control are integrated. Perceptual systems provide basic support data for autonomous or auxiliary driving systems. The effectiveness of the system becomes the basic guarantee of the effectiveness of the whole autopilot system. This paper designs an object detector model based on the deep convolution neural network. The object detection framework based on deep convolution neural network is divided into two research directions. One is to use the region recommendation method and the feature extraction of deep convolution neural network. The object detection framework is constructed by combining the methods. The other is to regress the region frame and object recognition directly from the deep convolution neural network. It also discusses how to train and test the performance of road scene object detector model under different methods. The road scene object detector designed in this paper uses the convolution layer of deep convolution neural network to extract the feature of image data surface. Then the region frame extraction algorithm is used to locate the specific position of the road scene object in the image. Finally, the detector model trained under the high performance computer is embedded in the embedded development system. The road scene object detection model is tested for road object detection under different light conditions during the day and at night, and good detection results are obtained. The pedestrian detection accuracy in the road scene is 83%, and for cars and buses in the road scene, Motorcycle, bicycle and other vehicles, the average accuracy of the detection is 81.
【学位授予单位】：北方工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【参考文献】