摘要
使用卷积神经网络在对视频流进行连续目标检测时,因光照和角度等环境的不确定因素会出现某一帧漏检、错检或多帧连续漏检的问题。针对这一问题,基于视频流中的时间相关性,提出一种基于时间相关性的置信度矫正算法,以显著降低目标漏检和错检率。该算法能够针对异常检测数据,使用指数平滑法对置信度较低的漏检目标进行预测矫正;对错检数据的置信度进行抑制。通过多个真实数据集的验证,结果表明,通过置信度矫正后的目标检测性能得到显著的提高,MAP平均提高了7.7%。
In the process of using convolutional neural networks for continuous target detection in a video stream, due to environmental uncertainties such as illumination and angle, there will be problems such as missed detection of a certain frame, wrong detection, or continuous missed detection of multiple frames. To solve this problem, based on the time correlation in the video stream, a confidence correction algorithm based on the time correlation is proposed to significantly reduce the target missed detection and false detection rate. The algorithm can use exponential smoothing method to predict and correct missed targets with low confidence for abnormal detection data;suppress the confidence of misdetected data. Through the verification of multiple real data sets, the results show that the target detection performance after the confidence correction is significantly improved, and the MAP is increased by 7.7% on average.
作者
陈颖
焦良葆
曹雪虹
Chen Ying;Jiao Liangbao;Cao Xuehong(School of Communication and Information Engineering,Nanjing University of Posts and Telecommunications,Nanjing 210003;Artificial Intelligence Industry Technology Research Institute,Nanjing Institute of Technology,Naning 211167)
出处
《现代计算机》
2021年第33期21-26,共6页
Modern Computer
关键词
目标检测
时间相关性
漏检
错检
置信度矫正
target detection
time correlation
missed detection
wrong detection
confidence correction
