Research On Indoor Obstacle Avoidance Platform Based On NVIDIA Jetson TX1

Indoor mobile robots have a wide range of applications in military,industrial and civil applications.In order to realize the basic functions of indoor mobile robots for autonomous navigation and obstacle avoidance,a lidar sensor is usually used for environment sensing.However,due to the working mechanism of laser radar sensor,although the position of the obstacle can be determined,but the identification and tracking of the obstacle cannot be realized.Human bodies are the most common moving obstacles indoors.And human movements usually have a certain direction,but also with diversification.Thus If they are treated like other ordinary obstacles,the navigation decision will probably be invalid.At the same time,with the leap of computing power and with the help of visual images,robot can effectively identify and track objects.Therefore,how to fuse laser radar and image data to sense the environment is now a research hot spot in the field of mobile robots.Aiming at the drawback that a single lidar sensor can not estimate the direction and speed of a moving person correctly,which will lead to the failure of navigation and obstacle avoidance,this paper brings out an indoor environment sensing method which relies on multi-sensor information fusion.Our method mainly focus on using RGB image data to obtain the human body segmentation mask with the image semantic segmentation algorithm based on deep neural network.And then we fuse with the lidar data with human segmentation mask to obtain the information about human positions.We then construct the social navigation layer to make the human body position information written into local cost map in ROS navigation stack.And finally we realize a navigation and obstacle avoidance method based on human body position.This paper mainly works on the following aspects:The image semantic segmentation algorithm based on deep learning.We first theoretically analysis the resource consumption of the modern basic neural network module on edge device.And find out that the ESPNet module shows better performance than any other module compared in this paper,thus its network structure is selected.We then retrain ESPNet and compare it with PSPNet,which is often used as baseline for semantic segmentation algorithm.And we then accelerate thesegmenting progress with Tensor RT frame and deploy the segmentation algorithm on TX1.We study on the navigation and collision avoidance based on robot operating system.We fuse the human body mask with lidar data obtained in the same period and then get the position of human through out calculation process.The construction of social navigation layer is then followed,which has a function of writing the human body cost into the local cost map.And finally we realize the navigation and obstacle avoidance on robot based on ROS frame.(3)We select the leading AI edge device,the NVIDIA Jetson TX1,which have a good performance on power consumption control and computing ability,and then design the hardware and software architecture of the system,and prove that our methods is workable and can be applied on the platform with limited computing power.