电子科技大学学报2025，Vol.54Issue(3)：401-410,10.DOI:10.12178/1001-0548.2024122
多特征融合的目标物体导航方法
Target object navigation method based on multi-feature fusion
摘要
Abstract
Target object navigation is the process of reaching the expected target object based on visual observation in an unknown environment.In this process,it is crucial to find the direction of the target object from visual observation.A target object navigation method based on multi-feature fusion is proposed to address this issue.Using a feature fusion module to fuse visual features that contain the overall and local information of the navigation environment and the text features that refer to the semantics of the target object,the method obtains the directional features that represent the navigation direction and environmental features of the navigation environment.And then this method associates the visual representation with the navigation direction,thereby guiding the generation of navigation actions,constraining the agent to navigate towards the direction of the target object,and improving the success rate and efficiency of the model's navigation.Experiments on the AI2 Thor dataset show that compared to the benchmark model,the navigation success rate SR has increased by 11.7 percentage points,and the navigation success path length weighted ratio SPL has increased by 0.093;Compared with current advanced methods,SR has increased by 2.1 percentage points and SPL has increased by 0.008.The experimental results have demonstrated the accuracy and efficiency of this method.关键词
目标物体导航/多特征融合/多头注意力机制/深度强化学习Key words
target object navigation/multi feature fusion/multi head attention mechanism/deep reinforcement learning分类
信息技术与安全科学引用本文复制引用
毕盛,林华伟,董敏..多特征融合的目标物体导航方法[J].电子科技大学学报,2025,54(3):401-410,10.基金项目
广东省科技计划(2020A0505100015) （2020A0505100015）
广东省高校教师特色创新研究项目(2022DZXX03) （2022DZXX03）
