Research On Human Object Detection Technology Based On Machine Learning

With the rapid development of artificial intelligence and computer vision technology,the smart algorithm in machine learning for object detection has become a research hotspot.The object detection is a kind of technology that can locate and classify the interested objects in the image.We started from the field of machine learning,analyzed the advantages and disadvantages of the traditional algorithm,and focused on the neural network algorithm.We compared state-of-the-art technologies based on the neural network and proposed a light-weight human object detection network for embedded environment,which solves the problem that network is difficult to deploy.The main contents and achievements of this thesis are as follows:(1)The feature extraction of images and classifier in traditional algorithms are introduced in this thesis.As a model of traditional object detection,the deformable component model becomes the basis of machine learning.With the improvement of deep learning and computer performance,the neural network has opened a new chapter for the field of target detection.In this thesis,the components and principles of neural network are explained in detail,and the classical object detection network is introduced.(2)As a data-driven technology,convolutional neural network has the characteristics of high design difficulty,large amount of data and a large amount of calculation,which is difficult to be applied in mobile devices with very limited computing resources.In the embedded environment,there is no high-speed network connection and high-performance computer.The CPU,power consumption and prices limit the application of large-scale networks.We studied this problem in detail,studied some lightweight backbone networks,and proposed two object detection networks suitable for embedded devices.(3)The M-YOLO and S-YOLO networks proposed in this thesis are lightweight CNNs based on the deep separable convolution,which is a non-standard convolution and is able to reduce the computation significantly.Two YOLO layers are used as the output in the network.The output of the network adopts multi-scale feature fusion,which improves the performance of the network.In this thesis,the standard datasets are used to evaluate the network.Our models are compared with the state-of-the-art models.