Real-time Detection Of Pedestrian And Vehicle Targets Based On UAV

Fast and accurate detection of pedestrians and vehicles in UAV images is a meaningful but extremely challenging task,which is widely used in military reconnaissance,traffic control,and rescue tasks in remote areas.However,because the drone is a small mobile device,its memory and computing power are very limited,which makes how to ensure the real-time detection process has always been a difficult problem.Taking into account the particularity of pedestrian and vehicle detection on UAV platforms,the purpose of this article is to design a lightweight real-time detection model that can be used in UAV multi-target detection tasks in order to better solve small target detection the problems that exist in the process,and then improve the application scenarios of drone technology,make more drone projects land in life.This paper analyzes the characteristics of small targets under drones,and constructs a small target dataset WIZ,which contains two types of targets,pedestrians and vehicles;for UAV images,the background is complex,the target is small,and the target scale changes greatly.The computing power of human-machine equipment is weak,which makes the direct use of mainstream target detection algorithms to solve the problem of low detection accuracy and large amount of calculation.Several improvements and optimizations have been made to the original SSD algorithm: First,the original SSD algorithm model parameter is too large.The problem is to reduce the number of convolutional layers and the number of detection layers.At the same time,according to the size and aspect ratio of the target to be detected in the UAV image,the appropriate anchor parameters are redesigned for each detection layer,so that each anchor can be more Good match to the target to be detected;secondly,in view of the problem that the SSD algorithm’s network structure has poor feature extraction capabilities for small targets,a lightweight receptive field enhancement module is proposed to improve the network’s ability to small targets while controlling the increase in network parameters as much as possible.The feature representation ability of;Finally,the use of spatial separable convolution and depth separable convolution to replace ordinary convolution for detection,which significantly improves the detection speed of the model,and the performance of the WIZ data set and mainstream detectors A more comprehensive comparative analysis was carried out.In summary,this paper constructs a small target data set WIZ for UAVs containing pedestrians and vehicles,improves and optimizes the original SSD algorithm,and proposes a model lightweight and lightweight receptive field enhancement module,and based on this the above uses spatially separable convolution and depth separable convolution to replace ordinary convolution to improve the detection speed.From the experimental performance of the above improvements on the WIZ data set,the above improvements have achieved good performance in small target detection tasks under UAVs and have good guiding significance for small target detection under UAVs.