基于红外与可见光图像融合的交通标志检测

针对大多数交通标志检测方法易受光照和天气影响的问题,提出一种基于红外与可见光融合的交通标志检测方法。首先,分别在红外与可见光图像中提取交通标志的形状和颜色特征,对目标进行粗定位,并将结果放入候选组内进行分类与整理;然后,将候选组放入卷积神经网络,对粗定位结果赋予相应的置信度;最后,根据置信度对红外与可见光的粗定位结果进行决策融合,并将融合结果展示在可见光图像中。实验结果表明,该算法对检测环境的变化更加鲁棒,在白天、夜晚和不良天气等多种环境下有效地提高了检测效率,具有良好的适用范围。

An object tracking and global localization method using the cylindrical projection of omnidirectional image

We present an omnidirectional vision system we have implemented to provide our mobile robot with a fast tracking and robust localization capability. An algorithm is proposed to do reconstruction of the environment from the omnidirectional image and global localization of the robot in the context of the Middle Size League RoboCup field. This is accomplished by learning a set of visual landmarks such as the goals and the corner posts. Due to the dynamic changing environment and the partially observable landmarks, four localization cases are discussed in order to get robust localization performance. Localization is performed using a method that matches the observed landmarks, i.e. color blobs, which are extracted from the environment. The advantages of the cylindrical projection are discussed giving special consideration to the characteristics of the visual landmark and the meaning of the blob extraction. The analysis is established based on real time experiments with our omnidirectional vision system and the actual mobile robot. The comparative studies are presented and the feasibility of the method is shown.