| The aging problem in modern society is becoming increasingly serious.Due to the expectation that robots can alleviate labor shortages,indoor mobile service robots have been developed and widely applied in production and daily life.Indoor mobile robots require the comprehensive application of various key technologies,such as target detection,autonomous control,positioning and navigation.Target detection is an important research area for indoor mobile service robots,which involves the ability of robots to recognize and track objects in their environment.However,when performing target detection in the working environment of indoor mobile service robots,the uneven lighting in the images captured by cameras seriously affects the accuracy of target detection.Moreover,the embedded platforms of indoor mobile service robots often suffer from insufficient computing resources,resulting in low accuracy and slow operation of target detection algorithms during deployment.These issues have limited the development and implementation of indoor mobile service robots.To address these problems,this thesis focuses on the following three research and improvement areas:(1)In response to the problem of uneven illumination in images captured by cameras,this thesis proposes an image correction algorithm based on a two-dimensional gamma function.The algorithm converts RGB images to the HSV color space,uses mean filtering to extract the illumination component in the brightness of the HSV color space,and corrects the illumination component using a two-dimensional gamma function according to the set brightness standard.The corrected illumination component is then used as the new brightness,and the hue and saturation are converted to the corrected new RGB image.Experimental results show that the algorithm significantly improves the standard deviation,average gradient,and entropy of images after correcting for uneven illumination.Additionally,the algorithm’s running efficiency completely meets the requirements of embedded devices.(2)In response to the problem of slow running speed of object detection algorithms due to insufficient computational resources on embedded devices,this thesis proposes a lightweight object detection algorithm,DID-YOLO,based on the YOLOv5 s model.The algorithm reconstructs the backbone network of YOLOv5 s and introduces depth-wise separable convolution and inverted residual structure.The depth-wise separable convolution has lower computation cost than regular convolution,while the inverted residual structure is used to compensate for the information loss caused by the introduction of depth-wise separable convolution in the backbone network.Experimental results show that the reconstructed DID-YOLO model has a size of only 3.63 MB,which is 48.65% smaller than the original network model.DID-YOLO achieves a real-time image processing speed of 31.2 frames per second on Jetson AGX Xavier.(3)This thesis proposes a feature compensating two-stage knowledge distillation algorithm and output layers to address the problem of decreased detection accuracy in DID-YOLO after the reconstruction of the backbone network.In this algorithm,the guidance learning of DID-YOLO’s feature layers is performed by introducing the output of the teacher network’s feature layers in addition to the output layer knowledge distillation.This effectively compensates for the information loss in DID-YOLO’s feature layers.Experimental results show that the distilled algorithm can improve the m AP@0.5 of DID-YOLO to 73.86%.This experiment proves that the proposed knowledge distillation algorithm can effectively improve the recognition accuracy of DID-YOLO without changing its network structure.Compared with similar mainstream algorithms,the DIDYOLO algorithm proposed in this thesis has higher accuracy,smaller model size,and faster processing speed in object detection tasks after knowledge distillation. |
PDF Full Text Request