A Multi-band Object Detection System Based On Objectness Enhancement Model

The intelligent drone becomes a hot research topic with well-developed drone technologies and deep learning.However,real-time scene parsing through object detection running on a UAV platform is very challenging,due to the limited memory and computing power of embedded devices.In recent years,mainstream object detection algorithms based on deep learning have two categories: Two-Stage algorithm and One-Stage algorithm.The two-stage algorithm has high detection accuracy but slow detection speed,while the one-stage algorithm has fast speed but lacks accuracy.This paper combines the advantages of these two algorithms and uses the object prior to enhance the detection effect of the One-Stage algorithm,and develops an airborne multi-band object detection system based on the objectness enhanced model.This system is to equip the intelligent drone with a coaxial device consisting of an infrared camera and a visible light camera for collecting images,and to apply embedded platform for all-day object detection and sending object detection results to the flight control system in real time.The system provides drones with visual perception capabilities,making them more intelligent.The main research work is as follows:(1)This paper introduces the idea of semantic segmentation to One-Stage object detection algorithm,designing an objectness estimation module to form the objectness enhanced object detection algorithm OEDet(Single-Shot Detector with Objectness Enhancement).The objectness estimation module estimates the objectness of the pixels in the full image,and outputs probability distribution for each type of object to provide pixel-level objectness priors for the object detection task and suppress the interference of background features on the detection results.(2)In this paper,a multi-scale fusion module is designed by using the dilated convolution.The multi-scale fusion module integrates feature information of different scales to enhance the multi-scale information of the shallow features of the One-Stage detection network to improve the accuracy of object detection.(3)This paper has developed an airborne system of multi-band object detection based on deep learning.Based on the OEDet and the model compression,this paper proposes a lightweight objectness enhanced object detection model OEDet-Lite to adapt to the characteristics of UAV aerial targets and the computing resources of the embedded platform.The major benchmarks including VOC and COCO are applied in experiments to test its effectiveness,the OEDet object detection model in this paper has a mAP of 81.7 on VOC2007-test and a mAP of 32.8 on COCO test-dev2018 with an input image size of 512 ×512.The OEDet are much more accurate than One-Stage algorithm and faster than Two-Stage algorithm,achieving the balance between object detection accuracy and speed.The One-Stage object detection algorithm SSD(Params:23.75 M,Flops:30.45 G,VIS mAP:88.06,IR mAP:90.03)has achieved significant improvement on the datasets of infrared and visible light vehicle from aerial view,when compared with mainstream the OEDet-Lite object detection model(Params:14.22 M,Flops:0.4712 G,VIS mAP:88.94,IR mAP:89.46).