本文基于IoU网络提出一种IT-AWCR(IoU network tracking with adaptive weighted characteristic responses)目标跟踪算法.首先,根据目标运动速度设计了目标搜索区域确定策略,通过理论分析使用ResNet50的Block3、Block4卷积块的输出分别作为目标的浅层和深层特征表示;其次,以目标定位准确度和滤波模型抗干扰能力为评价指标通过优化算法自适应计算目标深、浅特征响应加权权重,然后从加权融合响应中获取目标粗略位置和边界框,经扰动操作获取多个候选边界框输入IoU调制-预测网络预测IoU值,取最大IoU对应边界框为最终预测目标边界框;最后,根据训练样本的相关学习权重和样本间相似度更新生成样本集,基于样本集采用稀疏优化策略实现了滤波模型更新.OTB2015和VOT2018标准数据集上的实验结果验证了本文算法的有效性.
IT-AWCR target tracking algorithm based on the IoU network is proposed in this paper. Firstly, a determination strategy for the target searching area is designed according to the velocity of target, the outputs of Block3 and Block4 convolutional layers in ResNet50 is used as the shallow and deep feature representations of target respectively by means of theoretical analysis; Secondly, with performance indexes of the accuracy of target location and the anti-interference ability of the filter model, the weights of the target deep and shallow feature responses are computed adaptively through the optimization algorithm. The rough target position and bounding box are obtained by the weighted fusion response, and multiple candidate bounding boxes obtained by perturbation operation, they are entered the IoU modulation-prediction network to predict IoU values, taking the bounding box corresponding to the largest IoU as the final predicted target bounding box. Finally, the sample set is updated according to the relevant learning weights of the training samples and the similarities between this samples. Based on the sample set, the sparse optimization strategy is used to achieve the filter model update. The results of experiments on the OTB2015 and VOT2018 show that the effectiveness of the proposed algorithm.