Abstract: Object detection is a fundamental problem for mobile robots working under indoor environments. Object detection is affected by dynamic environmental changes, especially illumination variation. Indoor illumination change is analyzed so as to research how to identify illumination condition in the environment rapidly from image space feature, and then switch object detection mode according to illumination state which means either visual or laser sensor will be picked, together with feature selection ability provided by deep learning, to guarantee object detection peformance. First, by extracting statistic feautures on Y component in CIEXYZ space of the image together with some other features, the illumination state of environment when getting image is estimtaed rapidly. If the illuminaiton state is appropriate, the R-CNN algorithm will be adopted to finish rapid object detection in image space. In poor or excessive illumination state, the point cloud from 3D laser sensor is tranformed into depth image which is then used by R-CNN to implement object detection. The experimental results verify the validity of the algorithm proposed by this paper.

Key words: intelligent robot, object detection, deep learning, Region-based Convolutional Neural Networks（R-CNN） algorithm, depth map

摘要： 物体检测是工作于室内环境的移动机器人必须解决的问题。物体检测受到环境动态变化的影响，其中尤以光照变化的影响最为明显。分析室内环境中光照变化特点，研究如何通过提取图像空间特征快速识别环境中光照状况，并以光照识别结果控制物体检测模式切换，在不同光照状态下，自适应地选择使用图像传感器或者激光传感器数据，结合深度学习的特征选择能力，保证物体检测性能。机器人运行时，首先通过提取图像在CIEXYZ空间Y分量上的统计特征，并结合一些其他特征，实现快速地对图像拍摄时所处环境的光照状态进行估计；在光照适中的情况下，利用R-CNN算法结合移动机器人特点，实现在图像空间下的快速物体检测；在光照不足或过强时，先把三维激光传感器获取的点云转换成深度图像，再利用R-CNN算法实现物体检测。实验结果表明了所提出算法的有效性。

关键词: 智能机器人, 物体检测, 深度学习, R-CNN算法, 深度图像