A Research On Specific Object Detection Technology Based On Airborne System

Object detection has always attracted attention of reasearchers due to it plays a key role in many applications such as intelligent security and autonomous driving.As the computing power of embedded systems improves,it becomes possible to deploy object detection on airborne systems.UAVs(unmanned aerial vehicle)are fast and flexible,it will increase application value in intelligent security,military and other fields that combined with object detection technology.Rectent years,object detection technology based on deep convolutional neural network has improved rapidly in detection accuracy and speed.However,how to improve the read-time performance of the object detection algorithm on the airborne system while maintaining the accuracy is stiil a problem.At the same time,it also faces challenge such as motion blur and small targets.In order to solve the above problems,This thesis studies from two aspects of algorithm design and engineering implementation,attempt to design and implement real-time pedestrian target detection system in RK3399 embedded system.In terms of algorithm design:1.This thesis designs a lightweight network PMNet(Pyramid of Multi-mixconv Network),Compared with the Mobile Net V3-large network,PMNet speeds up by 7.3% when the top-1 accuracy of the Image Net2012 dataset is loss only 0.6 %,and the parameters were reduced by 14.8 %,Compared to Res Net-153 network,the parameters was reduced by 92.3%.We designed a object detection network based on PMNet that can detect both humans and faces.We also optimized the training process using head annotation information,five-point annotation information on the face,and data augmentation,OHEM and other methods to improve object detection accuracy.2.The attention mechanism is introduced into the object detection algorithm in this thesis,and Optimized attention loss function to improve recall?Compared with ordinary FPN detection networks,On the Crowd Humam dataset,FPN detection networks with attention modules increased m AP by 16.3 % from 29.36 % to 34.14%,and increased by 3.8 % from 77.14 % to 80.09 % on UAV123 dataset.After introducing the attention mechanism,the detection accuracy of the PMNet-based detection network in the Crowd Human dataset and the UAV123 dataset was improved from 28.12 % to 31.56 % by 12.2 % and from 7.6.61 % to 79.45 % by 3.7In terms of engineering implementation:3.we use weight quantization technology to optimize the PMNet-based detection network in this thesis.meanwhile,we design and impolement the object detection scheme base on the airborne embedded system,and the average delay is reduced to 35.4ms in the RK3399 embedded environment.In this thesis,a basic real-time object detection system is implemented with a small loss of accuracy.4.This thesis combines the KCF object tracking algorithm to further improve the realtime performance of the entire object detection system.And the average delay of the detection scheme in this thesis is reduced to less than 20 ms.