Motion blur processing method for visual SLAM system based on local residual blur discrimination network
Motion blur processing method for visual SLAM system based on local residual blur discrimination network
复制标题
DOI：
10.1007/s12206-022-0640-6
复制
发表时间：
2022-07
影响因子：
1.6
通讯作者：
Jiahao Chen;Yehu Shen;Qixin Zhu;Quansheng Jiang;Ou Xie;Jing Miao
中科院分区：
文献类型：
--
作者：
Jiahao Chen;Yehu Shen;Qixin Zhu;Quansheng Jiang;Ou Xie;Jing Miao
In visual simultaneous localization and mapping (vSLAM) systems, motion blur often leads to insufficient number of matched features, resulting in tracking failure. Existing solutions often tackle this problem by restoring sharp images from blurry ones. However, the computational costs are high, and the restored sharp images are usually distorted. The effect of blurry image sequences to vSLAM system is analyzed, and the relationships between feature matching and motion blur are acquired to deal with the above mentioned problems. A local residual motion blur discrimination network is proposed to detect images with motion blur efficiently. Motion blur recognition results are coupled with a vSLAM system so that the feature extraction process is guided by the results from the local residual motion blur discrimination network. The performance of the vSLAM system can be effectively enhanced when it is applied to sequences with motion blur. Experimental results on the Technische Universität München dataset show that the proposed algorithm increases the average tracking length by about 200 frames compared with the original method on some image sequences with violent motions. This algorithm effectively improves the stability and accuracy of the vSLAM system.
