基于ECO-HC的无人机指定行人稳定实时跟踪

Stable real-time tracking of specified pedestrian of UAV based on ECO-HC

针对无人机(UAV)指定行人跟踪中目标遮挡严重、尺度变化大的问题,提出一种基于无人机指定行人稳定实时跟踪算法,即ECO-HC跟踪算法。提取方向梯度直方图(f HOG)和颜色(CN)特征,利用高效卷积算子(ECO)跟踪算法中训练样本的内插模型使其f HOG特征和CN特征融合;利用分解卷积算子、紧凑训练样本模型和稀疏模型更新策略降低过拟合风险,减少算法复杂度,其响应分数最大值的位置即为目标中心位置;引入快速尺度滤波方法自适应得到目标尺寸最优估计,提高跟踪精度。实验选取60段公开行人数据库,结果均优于近年其他4种优秀跟踪算法,其中平均距离精度(mean DP)和平均重叠率精度(mean OP)分别为83. 2%和87. 3%,平均速度为56. 3帧/s。对于无人机拍摄的自建行人数据库,亦具有良好的鲁棒性和实时的处理速度。

Aiming at the problem that the target occlusion is serious and the scale changes greatly in the pedestrian tracking of the UAV,a stable real-time tracking algorithm based on the designated pedestrians of the UAV is proposed which is ECO-HC tracking algorithm. The histogram of oriented gradient（ f HOG） and color name（ CN） features are extracted,and the FHOG feature and the CN feature are merged by using the interpolation model of the training samples in efficient convolution operator（ ECO） tracking algorithm. Using decomposition convolution operator,compact training sample model and sparse model updating strategy to reduce risk of over-fitting and reduce complexity of the algorithm. The position of the maximum value of the response fraction is the target center position. Fast scale filtering method is introduced to adaptively obtain the optimal estimation of the target size and improve the tracking precision. The experiment selects 60 segment public pedestrian databases,and the results are better than the other four outstanding tracking algorithms in recent years. Among them,the average distance precision（ mean DP） and average overlap precision（ mean OP） are 83. 2 % and 87. 3 %. The average speed is 56.3 frames per second. For the self built pedestrian database captured by UAV,it also has good robustness and real time processing speed.