Tiny Face Detection Technology Research For Mobile Robot

With the increasingly extensive application of mobile robots,people have higher demand for the function of robot face detection,especially for the remote micro face detection,which is of great significance to the application of mobile robots such as search and rescue,security and human-computer interaction.In this thesis,aiming at the detection task of small face target of mobile robot,the problem of image motion blur caused by robot movement,the detection problem of small target,and the lightweight deployment of deep neural network on mobile robot are studied.In order to better realize the task of face target detection when the robot is moving,it is necessary to solve the problem of image motion blur when the robot is moving.Based on wiener filtering algorithm,this thesis estimates the direction and scale information of motion blur using frequency domain method and space domain method respectively for the parameters of degradation function model used by the algorithm.The motion blur image with unknown parameters can be restored accurately.Through the experiments on the images actually collected by the robot,the desired effect can be achieved.After comparing several target detection algorithms,this thesis adopts YOLOv3 neural network model as target detection algorithm.In the field of target detection,methods based on deep learning have been proved to be more robust and more accurate than traditional detection methods in light intensity and environmental changes.In this thesis,the characteristics of micro-face target and the difficulty of micro-face target detection are analyzed,and the YOLOv3 neural network model is improved according to the conclusion of the analysis.The single-scale image input is interpolated and upsampled to double scale image pyramid input,which increases the available pixel information of small targets and increases the detection success rate of small targets.The prediction box of the back end is processed and gaussian penalty term is added to improve the accuracy of prediction of small target position on the image to a certain extent,and improve the accuracy and recall rate of YOLOv3 algorithm.At the same time,the lightweight deployment of YOLOv3 algorithm on mobile robots needs to be solved.As a typical representative of edge devices,mobile robots cannot deploy deep neural networks with large capacity and large computation like YOLOv3.In this thesis,the network structure is analyzed and channel pruning is carried out on YOLOv3 neural network.By using the parameters of BN layer and calculating a characteristic error for each channel,each channel can form an importance scale.The importance of each channel is ranked according to the importance scale,and then the unimportant channels are pruned.The structure of YOLOv3 neural network is simplified so that the algorithm can be successfully deployed on mobile robots.Finally,experiments show that the number of neural network parameters completed by pruning is greatly reduced,and the detection effect of small target face can be well guaranteed,and the application of practical mobile robots can be completed.In this thesis,the algorithm is deployed on the mobile robot to verify the rationality and performance of the algorithm.Different environments are set up and the results show that the algorithm can accurately detect the face target when the robot is moving.