基于稀疏约束深度学习的交通目标跟踪

Traffic Objects Tracking Algorithm Based on Deep Learning with Sparse Constraints

针对车辆视觉跟踪过程中普遍存在背景复杂、光线变化、尺度旋转等干扰,而现有跟踪算法抗扰动能力差、鲁棒性低的问题,构造了一种基于稀疏约束及深度学习的车辆目标跟踪算法,采用去噪自编码神经网络对包含正负样本的训练集进行特征提取。在正向传播过程中对隐层进行稀疏约束,反向传播微调阶段,对连接矩阵进行权值衰减的稀疏调整,增加神经网络的鲁棒性,实现神经网络不同隐层特征的高效提取,将网络的输出作为Logistics分类器的输入,学习获得车辆分类器,并采用粒子滤波在线跟踪目标。试验结果表明:对连接矩阵和隐层进行稀疏约束的去噪自编码神经网络具有较高的跟踪精度和较强的跟踪鲁棒性,在场景光照剧烈变化、车辆发生遮挡、三维旋转、尺度变化及快速移动时都能进行较好的跟踪,平均中心位置误差远小于对比方法,仅为2.3像素;而增量式学习(IVT)跟踪、在线自适应增强(OAB)跟踪、多示例学习(MIL)跟踪算法的平均中心位置误差分别为17.52像素、28.76像素和17.66像素;该方法的平均重叠率达83%,较IVT跟踪、MIL跟踪和OAB跟踪算法分别提高24.5%、42.2%、28.8%,满足智能交通监控的实际需求。

Aiming at address the disturbances of complex background,illumination changes,scale changes and rotation commonly existed in vehicle＇s visual tracking process and the problems of poor anti-disturbance ability and low robustness in most existing tracking algorithms,a vehicle target tracking method based on sparse constraint and deep learning was constructed.Stacked denoising autoencoder was used to extract feature for training sets which contained positive and negative samples.In the forward propagation process,the method of sparsity constraints was adopted for hidden layers and sparsity adjustment with the weight decay strategy was done for the connection matrix in the back-propagation fine-tuning phrase,which improved the robustness of neural networks and achieved efficient features extraction for different hidden layers.Then taking the outputs of network as the logistic classification layer＇s inputs,vehicle classifier was obtained by learning.Finally,particle filter algorithm was used to track online targets.The results show that the neural network based on stacked denoising autoencoder with sparsity constraints for thehidden layers and connection matrix has higher tracking accuracy and stronger robustness,which traces well in the conditions of dramatic illumination changes,vehicle occlusion,threedimensional rotation,scale changes and fast-moving.The average error on center position is only2.3pixel,far smaller than other tracking methods,while those of IVT tracker,OAB tracker and MIL tracker are 17.52 pixel,28.76 pixel,and 17.66 pixel,respectively.The average overlap rate of the method reaches 83%,higher than those of IVT tracker,MIL tracker,OAB tracker which rise 24.5%,42.2%,and 28.8%,respectively.Therefore it meets the practical need of intelligent traffic surveillance.