一种基于SSD的行人检测改进算法

An improved pedestrian detection algorithm based on SSD

在智能视频监控、自动驾驶、智能机器人等新兴领域中,行人检测作为其中最基本的一环起着关键性作用。由于在复杂环境下行人检测容易受到客观因素的影响,例如行人的成像尺寸差异大、相似物体多等造成行人检测精度低的问题,针对上述问题,对由卷积神经网络构成的SSD算法进行改进。为了提高基础卷积网络提取特征的能力,设计以ResNet101为基础的改进网络替换原来的VGGNet16基础网络;同时为了解决原SSD算法的浅层网络对于目标特征提取不准确的问题,增加了FPN网络对提取的浅层与深层卷积的特征进行了融合。改进的算法在行人数据集上进行训练和测试,实验结果表明,改进的SSD算法检测行人目标的平均精度比原SSD算法提高了6%,并能准确识别复杂场景中不同尺寸的行人目标,证明了改进算法的可行性。

Pedestrian detection plays a key role as one of the most basic parts in the emerging fields of intelligent video monitoring,automatic driving and intelligent robot.Because pedestrian detection is easily affected by objective factors in complex environment,such as the large difference of pedestrian imaging size and many similar objects,the detection accuracy of pedestrian is low.Aiming at the above problems,the SSD(single shot multibox detector)algorithm composed of convolutional neural network is improved.In order to improve the ability of extracting features of the basic convolution network,an improved network based on ResNet101 is designed to replace the original VGGNet16 basic network;In order to solve the problem of the inaccurate extraction of object features in the shallow network of the original SSD algorithm,the FPN network is added to fuse the features of the shallow and deep convolution.The improved algorithm is trained and tested on the pedestrian data set.The experimental results show that the average accuracy of the improved SSD algorithm is 6%higher than the original one,and the pedestrians of different sizes in complex scenes can be accurately identified,verifying the feasibility of the improved algorithm.