Pedestrian Detection Based On The UAV And Deep Learning

Object detection and recognition is a very important and basic supporting part in the field of computer vision,and it is also a kind of key linked with the field of robotics to realize interactive and autonomous operation of the environment.With the rise of hardware computing resources,deep learning methods in the field of machine learning come to show its advantages,which rely on building deep neural network based on a large amount of data to express the characteristics of picture information.This kind of method is more in line with the nature of the features.Compared with traditional methods,it has powerful generalization capabilities and higher performance.Although deep learning has achieved impressive results and attracted the attention of the whole society,it is limited by the impact of the limited computing power of small and medium-sized robot embedded devices and the relatively complex working environment status.With a large amount of computing resources,better performance can be achieved,but few applications are in the field of embedded devices.The purpose of this research is to study the topic of target detection based on deep learning,focus on the specific data characteristics of the perspective of the drone at the application background of the quadrotor drone in autonomous movement,to realize the real-time pedestrian detection on UAV.Firstly,the pedestrian data is collected by the drone in different environments.According to the format of the public dataset of PASCAL VOC,the image tagging tool is used to manually make the pedestrian detection training data and test data sets under the perspective of the drone,which are used to train the neural network and evaluate performance of model.Then,we choose the anchor-free object detection network frame CenterNet,which is faster than either the region proposal strategy or regression strategy,in order to achieve the requirement of the real-time detection in the UVA experiments.On this basis,a deep neural network with a main network architecture of ResNet-18 is constructed.In order to further improve the model's operating speed and reduce the number of model parameters,a deep separable convolution block is introduced to compress the model.A pedestrian target detection model is obtained from the collected pedestrian data training set.Finally,a large number of UAV deployment models were tested experimentally,and the actual performance of the model was tested under different scenarios and different numbers of targets.Due to the designed lightweight neural network,it can basically meet the design requirements of accuracy and real-time performance under the processing of onboard computing resources.