Research On Object Detection For Spherical Robot Based On Deep Learning

Spherical robot is a special mobile robot with spherical shape and rolling motion.Perception of the environment is the basis of mobile robots.However,the limitation of the volume and power of spherical robots leads to a shortage of computing resources,which puts forward high requirements on the performance of object detection algorithms.In addition,the instability of the ball and the way of rolling motion cause the spherical robot to shake more seriously.Furthermore,cost constraints,occlusion of the necessary mechanical structure of the robot,low viewing angle,and backlight cause poor image quality.Aiming at the above difficulties,this paper designs a lightweight object detection algorithm for the spherical robot visual perception system and a object detection system based on multi-sensor fusion.This paper verifies the effectiveness of the algorithm on the collected spherical robot object detection data set.The main research work and results of this article include:1.In view of the limited computing resources of spherical robots,a lightweight object detection algorithm suitable for spherical robots is designed.The specific structure of the algorithm is designed based on the principle of lightweight,and the feature extraction network part is designed based on the MobileNetV3 algorithm.The experimental verification shows that the object detection algorithm designed in this paper has excellent detection effect.Experimental verification on the spherical robot computing platform shows that the CPU and memory occupancy rate and real-time performance of the algorithm have reached the corresponding performance requirements.2.Aiming at the challenge of pure visual object detection for spherical robots,a spherical robot object detection system based on multi-sensor fusion is designed.The system feature extraction network integrates single-line lidar data,and uses inertial measurement unit data to correct image and radar data.Experiments compared with the pure visual object detection algorithm designed in this paper,the object detection network fused with single-line lidar data is better with a little increasing of the calculation.On the spherical robot object detection data set,the mAP of the object detection system is increased by 4.5%compared with the pure visual object detection algorithm,and it has a better detection result.3.In order to better verify the effectiveness of the algorithm,after comparing different algorithms and data sets,according to the characteristics and distribution of the spherical robot and its sensors,a collection method of the spherical robot multi-sensor target detection data set is designed.After independent calibration and joint calibration of the sensor,a total of 1661 images and their corresponding single-line lidar and inertial measurement unit data were collected in different scenarios,and the effectiveness of the paper's algorithm was verified by joint other public data sets.