基于帧图像语义上下文的地形推理策略  

Terrain Inference Strategy Based on Semantic Context of Frame Images

作　　者：王明军[1] 周俊[2] 屠珺[3] 

机构地区：[1]宁波工程学院,浙江宁波315016 [2]南京农业大学,江苏南京210031 [3]上海交通大学,上海200240

出　　处：《机器人》2012年第3期299-306,共8页Robot

基　　金：国家自然科学基金资助项目(51075217;31071325);宁波市自然科学基金资助项目(2011A610198).

摘　　要：基于场景帧图像语义上下文概念,提出一种融合不同场景时空信息的自适应在线地形推理策略TIBSC(Terrain Inference Based on Semantic Context).首先,提取不同场景图像近视场像素的特征向量和地形类别,构建地形样本候选数据库;然后,从地形样本候选数据库中选取与当前场景图像远视场区域语义上下文最相似的若干帧地形样本,构建当前场景地形判别数据库;最后,基于当前场景地形判别数据库及贝叶斯法则,对当前场景远视场区域像素所属地形类别进行推理.基于数据集的仿真实验表明:在影响TIBSC推理精度的所有因素中,基于语义距离准则的最优样本选择以及样本数据的在线扩充两大模型变量起主导作用.同时,基于相同分类特征的TIBSC策略在推理精度上优于同类研究结果.Based on the concept of semantic context of flame images, an adaptive online terrain inference strategy （TIB SC, terrain inference based on semantic） is proposed, which incorporates spatiotemporal information of different scenes. Firstly, feature vectors and terrain categories of close-field-of-view pixels of different scene images are extracted to construct the terrain sample candidate database. Secondly, samples with most similar semantic context to that of the distant-field-of-view region of current scene are selected from the terrain sample candidate database to further construct the terrain discriminant database of the current scene. Finally, based on the discriminant database of the current scene and the Bayesian rule, terrain categories of distant-field-of-view pixels of the current scene are inferred. Results of the simulation experiments based on database show that the optimal sample selection based on semantic distance criterion and the online sample expansion play the dominant role among all the factors influencing the inference accuracy of TIBSC. And the results also indicate that, TIBSC model outperforms other existing methods in the term of inference accuracy.

关 键 词：在线地形推理 场景分析 语义上下文 

分 类 号：TP24[自动化与计算机技术—检测技术与自动化装置]