| With the significant improvement of computer computing performance and the development of artificial intelligence technology in various fields,artificial intelligence products gradually appear in people's vision,such as mobile robots and driverless cars.The environment perception of mobile robot means that the robot realizes the perceives and understands the surrounding environment through a series of self-configured sensors,such as radar and camera,which is the premise and basis of decision-making and control.Visual perception technology has become a hot research because of its low hardware cost and large amount of perceived information.With the rapid development of deep learning in image processing in recent years,valuable research results continue to emerge in the sub-problems of image classification,obj ect detection and semantic segmentation,which provide new solutions for vision-based environmental perception.This paper studies the application of visual environment perception technology based on deep learning,and presents a unified algorithm architecture for object detection and road segmentation,which can simultaneously detect targets and segment drivable roads,so as to guide the target tracking and obstacle avoidance of mobile robots.The specific works of the paper are as follows:1.In terms of detection algorithms,this thesis studies two mainstream methods in the field of deep learning based object detection,namely,the method based on candidate boxes and based on direct regression.A target detection algorithm combining ROI and YOLO algorithm is proposed which can improve the accuracy of the algorithm while ensuring the timeliness of operation.In terms of semantic segmentation,this thesis analyzes the principle and compares the performance of current representative algorithms,and finally adopts the semantic segmentation algorithm based on FCN as the segmentation model.2.Aiming at the real-time processing of environment information of mobile robots,this thesis designs a joint architecture of detection and segmentation,which shares the same feature extraction network and improves the performance of the two sub-tasks to some extent while reducing reasoning time.3.In order to test the performance of the detection and segmentation joint algorithm,the model is trained and evaluated on the open data set.In order to expand the applicable environment of the algorithm,a target detection and semantic segmentation data set of indoor corridor environment is constructed according to the standard of open data set.4.This thesis designs and builds the hardware platform of vision-based environment perception system of mobile robot,which consists of wheeled robot and its carrying cameras and a portable computer.This thesis also designs a simple motion control strategy.According to the perceived environmental information,the mobile robot can move along the drivable area in both indoor and outdoor environments,and can complete other extensibility tasks according to the detection results of specific targets.To sum up,this thesis studies and optimizes the algorithm of target detection and semantic segmentation,establishes a joint architecture of detection and segmentation through sharing feature extraction network,and builds a hardware platform of mobile robot to carry out practical application experiments. |
