完美论文网提供代写毕业论文、代写本科论文、代写硕士论文、论文发表及查重服务!QQ:505676555

电话:18796337551
当前位置:完美论文网信息系统 → 文章正文

基于深度学习的视频目标检测算法研究

作者:完美论文网  来源:www.wmlunwen.com  发布时间:2019/10/10 8:45:30  

摘要:基于视觉的目标检测是图像处理、计算机视觉、模式识别等众多学科的交叉研究课题,在视频监控、自主导航等领域,具有重要的理论研究意义和实际应用价值。目标检测分为静态图片目标检测和动态视频目标检测,对于静态图片,主要考虑的是目标检测的精度问题,影响因素主要有:图片中目标重叠率高、图片中的目标占图比小、图像分辨率低等。但随着相关课题的深入研究,目前对于静态图片中目标检测的准确率已很高,研究者都开始运用深度学习来进行动态即视频目标检测。对于视频目标检测需要考虑的不仅仅是精确率,还需考虑实时性的问题。视频目标检测往往存在如下困难:数据集数量少、在构建训练样本时出现难易样本不平衡、视频中远处的目标或拥挤的检测效果不佳、视频中目标检测的精确率和实时性很难都达到最佳等问题。针对上述问题,本文在现有算法的基础上,进一步研究了基于深度学习的视频目标检测算法,主要内容包括:

1. 通过车载摄像头与道路摄像头等采集包含车辆与行人目标的视频,并对采集视频进行处理赋予目标标签,建立车辆与行人数据集。

2. 研究与分析目前已有的经典视频目标检测算法。针对Faster RCNN的实时性不够、远处和近处场景中的目标车辆所占图片比差异较大影响检测效果等缺点,提出了一种基于Faster RCNN框架的车辆目标检测改进算法,该算法通过采用k-means算法获得合适的anchors长宽比,改进RPN网络结构,结合多尺度训练等来改善系统的性能。实验结果表明该改进确实能够很好的提高准确率和加快速度。

3. 针对YOLOv2的检测精度低于Faster RCNN、数据集中的难样本的识别率不高等问题,提出一种以YOLOv2模型为基础的改进算法。该算法引入了新的损失函数,增加检测窗口内细胞单元的数量,并改进模型中anchors的数量与大小,以此设计了自动学习车辆特征,实现高精度与快速的车辆自动检测系统。通过实验发现该改进确实在不降低速度的情况下很大程度上改善了精度。

4. 目前单独基于视频目标检测算法的应用有一定的缺陷,不利于实际应用。因此,本文将行人目标检测与行人重识别相结合,提出了一种基于多目标的跨摄像头视频目标检测算法。首先,根据公开的行人数据集训练行人目标检测算法,再将感兴趣的行人与行人重识别库中的行人进行比对,训练行人重识别算法,需要判断感兴趣的行人是否在待预测的视频之中出现,如果出现的话则返回行人的全身位置信息,视频号,帧号以及对应的身份id。实验结果表明,该算法具有一定的实用价值。

Vision-based object detection is across-disciplinary research subject of image processing, computer vision,pattern recognition, etc. It has important theoretical research significanceand practical application value in video surveillance, autonomous navigationand other fields. Object detection is divided into static image objectdetection and dynamic video object detection. For static images, the mainconsideration is the accuracy of object detection. The main influencing factorsinclude: high object overlap rate in the image, small object proportion in theimage, low image resolution and so on. However, with the further study ofrelated subjects, the accuracy of object detection in static images has beenvery high at present, and researchers have begun to use deep learning to carryout dynamic detection, that is, video object detection. For video objectdetection, not only precision but also real-time performance should beconsidered. Video object detection often has the following problems: a smallnumber of data sets, imbalance of hard and easy samples in the construction oftraining samples, poor detection effect of distant objects or crowding invideo, and difficulty in achieving the best precision and real-time performanceof object detection in video. Based on the existing algorithms, this thesisfurther studies the video object detection algorithm based on deep learning.The main contents include:

1.     Videocontaining vehicle and pedestrian objects is collected through vehicle camerasand road cameras, and the object tag is assigned to the collected video toestablish vehicle and pedestrian data sets.

2.     theexisting classical video object detection algorithms are studied and analyzed.Aiming at the lack of real-time performance of Faster RCNN and the largedifference in the image proportion of vehicles in distant and near scenes, animproved object detection algorithm based on Faster RCNN is proposed. Thealgorithm by using k-means algorithm to obtain appropriate anchors aspectratio, improves the RPN network structure, and combines with multi-scaletraining to improve the performance of the system. The experimental resultsshow that the improvement can improve the precision and speed very well.

3.     Animproved algorithm based on YOLOv2 is proposed to solve the problem that thedetection accuracy of YOLOv2 is lower than that of Faster RCNN and therecognition rate of difficult samples in the dataset is not high. The algorithmintroduces a new loss function, increases the number of cell units in thedetection window, and improves the number and size of anchors in the model, soas to design an automatic learning vehicle features to realize the highprecision and rapid vehicle automatic detection system. It has beenexperimentally found that this improvement does improve the precision greatlywithout reducing the speed.

4.     Atpresent, the application based on the video object detection algorithm alonehas certain defects, which is not conducive to practical applications.Therefore, this thesis combines pedestrian detection with personre-identification, and proposes a multi-object cross-camera video objectdetection algorithm. Firstly, pedestrian detection algorithm was trainedaccording to the pedestrian dataset, then the interest of pedestrians arecompared with the pedestrians in the person re-identification library, and theperson re-identification algorithm is trained to determine whether theinterested pedestrian is in the videos to be predicted. If the pedestrianappears in the video, the algorithm returns the pedestrian's position, thenumber of the video, the number of the frame, and the id of the correspondingidentity. The experimental results show that the algorithm has certainpractical value.

关键词:实时目标检测;FasterRCNN;YOLOv2;k-means算法;区域提案网络;损失函数

real-time object detection; Faster RCNN;YOLOv2; K-means; regional proposal network; loss function

联系方式

客服QQ 505676555
客服热线18796337551
网站地址 www.wmlunwen.com
郑重承诺 原创,包修改,包通过!
完美论文网真诚欢迎新老客户的光临与惠顾!