基于FasterMDNet的视频目标跟踪算法

Video Target Tracking Algorithm Using FasterMDNet

多域卷积神经网络(MDNet)算法在卷积层采用选择性搜索的方式来提取候选框,因此它没有共享完整图像特征,从而导致在线视频目标跟踪速度慢。针对这个问题,提出一种快速多域卷积神经网络(FasterMDNet)视频目标跟踪算法。FasterMDNet是建立在MDNet基础上的一种模型,在卷积层后面引入RPN(Region Proposal Network)网络,优化了损失函数,共享完整图像卷积特征,加快候选区域建议框(ROI)更高效的生成;为了更好地获得目标和背景信息表示,在RPN网络后加入ROIAlign层,对提取的候选区域建议框特征图用双线性插值方法来提高感受野的分辨率。该算法对目标跟踪基准数据集OTB2013、OTB2015、VOT2016进行了评估,并与前沿的跟踪算法做对比,实验结果证明,该算法跟踪准确率优于其他对比方法,并且对比相同实验环境下MDNet算法,在线跟踪速度提高了近12倍。

Multi-Domain convolutional neural Network(MDNet)algorithm uses selective search in the convolutional layer to extract candidate boxes,therefore,it cannot share full image features,resulting in slow tracking of online video target.A Faster Multi-Domain convolutional neural Network(FasterMDNet)video target tracking algorithm is proposed to solve this problem.FasterMDNet is based on MDNet,it adds the RPN(Region Proposal Network)behind the convolutional layer,optimizes the loss function,shares the full image convolution feature,and accelerates the generation of candidate ROI(Region of Interest)more efficiently.In order to better obtain the target and background information representation,the ROIAlign layer is added after the RPN network,for the feature map of extracted candidate ROI,which uses bilinear interpolation to improve receptive field resolution.The proposed algorithm is evaluated in multiple target tracking benchmark datasets including OTB2013,OTB2015,and VOT2016,and compares with the state-of-the-art algorithm.Experimental results show that tracking performance of the proposed algorithm is better than other comparison methods,and the online tracking speed is nearly 12 times faster than MDNet algorithm in the same experimental environment.