基于深度降噪自编码器的多特征目标融合跟踪算法

Multi-feature target fusion tracking algorithm based on deep denoising self-encoder

针对在目标遮挡、光线变化、目标模糊等情况下的目标跟踪算法抗干扰能力较差的问题,提出了一种基于深度降噪自动编码器的多特征目标融合跟踪算法。该方法首先引入稳像和图像去雾算法以改善训练集数据和测试集数据的质量;再构建多特征深度降噪自动编码网络,基于深度神经网络的强大学习能力提取目标的颜色特征和均匀模式纹理特征;将两种特征加权融合输入到逻辑回归分类器,获得置信分数,更有效地区分目标和背景。最后,采用粒子滤波算法对目标进行跟踪。实验结果表明,该方法能够更准确地对存在目标遮挡、光线变化、目标模糊等干扰问题的视频进行跟踪。与传统方法相比,该方法成功率在上述三个方面平均分别提升33.73%、9.73%和12.80%;与近年流行算法相比,该方法成功率平均达到90.16%,实时性平均达到49.37 fps。

In order to solve the problems of target occlusion,ray change,target blurring and poor anti-jamming ability of existing tracking algorithms in the process of target tracking,We propose a multi-feature target fusion tracking algorithm based on depth denoising automatic encoder.At first,the quality of training set data and test data was improved by introducing algorithms of image stabilization and image defogging.Next,a multi-feature deep denoising automatic coding network is constructed to extract HSV color features and Uniform LBP texture features of the target based on the powerful learning ability of the deep neural network.Then,the two features are weighted and integrated into the logistic regression classifier to obtain the confidence score,which can effectively distinguish the target from the background.Finally,we used particle filter algorithm to track the target.The experimental results showed that this method can track the video with the interference problems such as target occlusion,ray change and target blurring more accurately.Compared with the traditional method,the accuracy of the method increased by an average of 33.73%,9.73%and 12.80%in the above three aspects;Compared with the popular algorithms in recent years,the method has an accuracy of 90.16%on average and a real-time average of 49.37 fps.