Research On Object Detection Based On Improved Faster RCNN Algorithm

In recent years,with the rapid development of drone technology,it has been widely used in military,commercial,and household fields,and is rapidly entering various industries.When executing tasks,it is necessary to have functions such as autonomous planning and target recognition,and to ensure timeliness and real-time detection of targets.Given the fast and better robustness of the Faster RCNN algorithm,it is of great significance to study the application of the improved Faster RCNN algorithm in object detection.This article mainly studies and improves the Faster RCNN object detection algorithm,and integrates it into vehicle object detection tasks from the perspective of unmanned aerial vehicles.The details are as follows:Firstly,this article selects the Visdragon dataset and preprocesses it for training and verifying the performance of the model.The dataset includes five vehicle categories: car,truck,van,bus,and freight＿car.Secondly,in the process of detecting small targets from the perspective of drones,the traditional Faster RCNN backbone network VGG16 has problems such as a large number of training feature parameters,poor performance,and difficulty in parameter adjustment,resulting in information loss.An improved algorithm based on two-stage network is proposed,which introduces HRNet network into the backbone network.By adopting the principle of high and low resolution sharing,high resolution always exists,enhancing information expression ability,and making it more suitable for small target detection.On the other hand,due to the mismatch between the target box generated from the BEV perspective and the image edges,it is easy to cause false positives and missed detections during the detection process.This article introduces the method of rotating object detection and introduces rotation variables in the RPN network to generate tilted candidate boxes that fit the edge of the image at any angle,effectively improving the accuracy of object detection and solving the problems of false detection and missed detection.Finally,to address the issue of improved detection accuracy but reduced speed in the improved algorithm.Propose a method for lightweight processing of network structures,namely lightweight backbone network structures.The backbone network after the improved algorithm has a deep level,a lot of Semantic information is extracted,and the computing speed is relatively slow.This paper introduces the peleenet network as the backbone network after the improved algorithm,so that the convolution layer and BN layer can be combined for computing at the same time,thus solving the problem of slow computing speed,and ensuring that the detection accuracy is improved by about 5.89% on the premise that the detection speed is unchanged.Verified the effectiveness of the improved Faster RCNN algorithm,providing a new approach for unmanned aerial vehicles to perform vehicle detection tasks.