Recognition Method And Application Based On Depth-CNN For The Robot Crossing Behavior

Indoor service robot require can replace humans independent walking, autonomous patrol function under the fire, hostage rescue, nuclear radiation and others dangerous circumstance. A key issue of indoor walking robot is how to identify the connection medium (such as doors, tunnels, caves) accurately and completed crossing based on the results of it in a structured environment. Currently, the infrared sensors, laser scanning, image edge detection methods are generally adopted in obstruction detection. For the following reasons leading to detection and identification often appear higher error rate:(1) Due to the complexity of the indoor environment, it is common with the appearance of the model edge feature scenes, and frequent shielding occurrence. If the framework model of the door is used for door identification, i.e., two vertical and one horizontal structure, it is lack of some hidden feature extraction and false detection rate is rather high with the occlusion disturbance especially.(2) There is not a lot of research on the posture of the doors but only existence of the door. Besides, seldom application based pure visual information has been used in the indoor navigation.In order to deal with the problems, This thesis propose a door identification method based depth convolution neural network (DCNN), which can improve the door recognition ability and crossing ability under the disturbance scenes. First, the automatic feature extraction methods are studied so as to avoid feature loss caused by modeling, and insufficient robustness of recognition due to the occurrence of occlusion with other methods. Second, the training data for DCNN are also generated and collected for modeling demand. To design the multi-gesture recognition scheme and used for door angle recognition. The relative position of the robot can be localized as the direction for door traversing. Then, based on the door images, the distance can be estimated for robot movement. Combined with the angel calculation, the door can be localized so the robot can perform passing through experiments. Finally, experiments are performed for sensitivity analysis, i.e. how much degree of occlusion can affect the door identification accuracy. It demonstrates that the DCNN has more advantage on door recognition with higher reliability and robustness.In this dissertation, on the one hand, the effective of door recognition model based on DCNN has been verified. On the other hand, the door crossing experiment has been completed by using the identification information of the door. This thesis can provide some reference for the establishment of a purely visual indoor navigation systems.