一种基于丰富视觉信息学习的3D场景物体标注算法

在智能服务机器人领域,根据场景的图像序列来完成场景重构及其中物体的检测与标注,是机器人场景理解、人机交互及后续服务的基础.本文针对基于RGB-D数据的3D场景物体标注展开研究,设计了一种充分融合颜色与深度数据的物体建模学习方法,并将其应用于图像目标快速检测及3D场景物体标注.离线学习阶段,在构建物体检测模型时,加入了物体颜色的高斯模型,与物体的RGB-D HOG特征一同构成先验模型.在线阶段,首先对待检测的场景图像进行超像素分割,将对场景图像的处理从以像素为单位转变为以超像素区域为单位;同时计算每个超像素区域的高斯颜色模型,并与物体的先验高斯颜色模型比对,筛选得到物体所在的候选超像素区域;然后,在筛选出的超像素周围进行滑动窗口搜索,计算目标物体出现在场景图像中的概率图;最后,将该概率图映射到3D场景中对应体素,联合该体素及其周边体素构建马尔科夫随机场,进而进行3D场景物体标注.实验结果表明,由于引入了超像素处理,以及基于颜色模型比对的超像素筛选,所处理的数据量大大减少,在基本不损失精度的同时,算法效率明显提高.

基于深度学习的移动物体检测

本文为解决深度学习中无法达到对多张照片移动目标物体识别的需求,提出一种基于深度学习的移动物体检测方法。首先利用前馈神经网络中目标定位技术设计目标定位模型,计算出目标边界框的大概位置,再基于时间空间融合方法提取特征,以3D卷积网络提取时间信息对移动目标行为进行预测,通过级联分类器剔除非有效目标,有效提高了移动物体检测的准确性和效率。

General moving objects recognition method based on graph embedding dimension reduction algorithm

Effective and robust recognition and tracking of objects are the key problems in visual surveillance systems. Most existing object recognition methods were designed with particular objects in mind. This study presents a general moving objects recognition method using global features of targets. Targets are extracted with an adaptive Gaussian mixture model and their silhouette images are captured and unified. A new objects silhouette database is built to provide abundant samples to train the subspace feature. This database is more convincing than the previous ones. A more effective dimension reduction method based on graph embedding is used to obtain the projection eigenvector. In our experiments, we show the effective performance of our method in addressing the moving objects recognition problem and its superiority compared with the previous methods.