基于IOU分析的稀疏视频检测技术研究

A novel sparse video detection method based on IOU estimation

基于深度学习的目标检测技术发展迅速,检测性能不断提高。然而,在视频检测应用中,由于视频数据量较大且实时性约束严格,导致目标检测算法的计算资源消耗极高。针对视频检测算法的巨额计算资源消耗问题,本文提出了一种基于深度学习的目标检测算法和目标追踪算法自适应结合的稀疏视频检测方法,能够动态地基于目标区域交并比(IOU)分析,自适应地利用计算资源消耗较小的目标追踪算法替代目标检测算法进行视频分析,从而在保障视频检测准确率的前提下,大幅降低计算资源开销,并进一步提高了视频检测的鲁棒性。

Recently,the performance of object detection algorithms based on deep learning is continually improved.However,in video detection applications,the computational resource consumption of object detection algorithms is more and more huge,which is caused by the processing speed constraint and data size of video.To address this,a sparse video detection method based on intersection over union(IOU)estimation is proposed.In the proposed method,object tracking algorithm with much lower computational resource consumption is adaptively activated by IOU estimation to replace object detection algorithm.Experimental result shows that the proposed method not only greatly reduces the overall computational resource consumption,but also improves robustness for video detection.