Research On Object Detection For Quadrotor Systems

Most of the inspection tasks are manual y inspected by workers in the current projects.However,there are a series of issues in manual inspection,such as,low efficiency,low inspection frequency,high cost and etc.Along with the UAV(Unmanned Aerial Vehicle)technology unceasing development,automatic inspection can largely replace the traditional manual inspection,which solves the current problems of manual inspection.Nowadays,the key subject of the current intelligent inspection research is the visual perception methods,therefore,the paper will be focused on object detection for quadrotor system,and there are two aspects in research programme:1.Research on onboard object detection method based on machine learning.Considering the limited computing power of onboard computers,the focus of research is on traditional machine learning methods.In this paper,the object detection method based on HOG(Histogram of Oriented Gridients)is improved.By using feature fusion strategy to strengthen features,the strategy is performed for HOG and LBP.Then input the designed features into SVM(Support Vector Machine)for object detection,so as to achieve the improvement of algorithm speed and accuracy.In the experiment,we also build the dataset,and trained the algorithm on the dataset,completed the algorithm evaluation finally.The experimental results show that the proposed algorithm achieves improved accuracy compared with the traditional algorithm of HOG + SVM.2.Research on off-board object detection method based on deep learning.Design an object detection scheme based on deep learning.The focus of the research is on the one-stage object detection model,which is designed to detect single-class targets in the UVA perspective under high-altitude inspection environment by designing a single-stage neural network.The model improved based on YOLO v3 took the adoption of 43-layer backbone network.And the object can be detected on two scales through feature fusion strategy.Besides,it also designed the loss function with the idea of Focal Loss.The experimental object of the article is excavator under the direct UAV view.We obtained the optimal solution of the model through the corresponding data augmentation in accordance with characteristics of the object after the completion of the preconstruction of the dataset.The test results showed that this model is more accurate and robust than YOLO v3 in the same input size,and has an increase of 10 FPS in frame number.