Open-air Target Detection And Counting Under UAV Platform

Open-air target detection and counting from the perspective of UAV is one of the important task in the field of artificial intelligence and computer vision in recent years.Deep convolutional neural network,feature fusion and other algorithms are mainly used to calculate the location,number and density of open-air targets in images.The development of computer vision has been able to meet the requirements of counting objects such as vehicles in aerial images and estimating crowd density in street images in recent years,however,due to the complex background and small targets of UAV aerial images,the simultaneous detection and counting of multiple targets such as people and vehicles in aerial images need to be further studied.This thesis takes the counting of open-air targets such as pedestrian and vehicles in aerial images of unmanned aerial vehicles as the research target.Firstly,an open-air target detection model based on improved YOLO v4 is designed to improve the detection and counting accuracy of small targets.Then an aerial crowd counting model based on deep convolutional network is designed to estimate the crowd congestion in the form of human body detection and density map generation.Finally,an open-air target detection and application system of UAV platform is designed to implement the algorithm mentioned above and optimize the algorithm at the same time to improve the system security,availability and stability.The innovation of this thesis is mainly reflected in the following three aspects:(1)For the detection and counting of open-air targets in aerial photographs,especially small targets,an open-air target detection method based on the improved YOLO v4 model is designed,which enhances the fusion of feature maps suitable for small target sensing fields,and introduces the attention mechanism and dilated convolution pool technology,so that the model can cover multiple scale targets.The detection and counting experiments is carried out on Vis Drone2021-DET dataset.The experimental results show that the open-air target detection method based on the improved YOLO v4 model can effectively combine the low-level location information and high-level feature information to improve the detection and counting accuracy.(2)A YOLO-CC network structure is proposed to improve the accuracy of crowd detection and counting in UAV images.Firstly,the human body is detected by using low-level position information through target detection algorithm.Then,multi-column convolutional neural network and dimensionality reduction are used to generate the crowd aggregation density map on the high-level feature map.The combination of target detection and density map generation can effectively improve the counting accuracy and strengthen the attention to small targets.Experiments on Vis Drone2021-COUNTING dataset show that the YOLO-CC model has a good effect on the detection and counting tasks of clusters of people.(3)The design and implementation of an unmanned aerial vehicle platform open-air target counting application system,which transplants the above network model to a cloud server,enables the application system to complete the automated processing and analysis of open-air target detection and counting,pedestrian detection and counting tasks.