Unmanned aerial vehicle ground detection method and system based on confidence

Abstract

The invention relates to an unmanned aerial vehicle ground detection method and system based on confidence, and the method comprises the steps: carrying out the forward reasoning of video data collected frame by frame based on a model file and a weight file of a trained target detection deep neural network, and obtaining a position region of a to-be-detected target, obtaining a candidate region corresponding to the position region of the to-be-detected target in the current video frame and a region consistent with the candidate region corresponding to the current video frame in the next videoframe, determining the position region of the to-be-detected target, taking the position region of the to-be-detected target as a target position, storing the target position in a target position setand carrying out outputting. According to the technical scheme provided by the invention, the video acquired by the unmanned aerial vehicle in real time is monitored by adopting the deep neural network target detection method, the detection efficiency of the unmanned aerial vehicle is improved, and the method and system have important application value.

Description

technical field [0001] The invention relates to the technical field of computer vision, in particular to a method and system for ground detection of an unmanned aerial vehicle based on confidence. Background technique [0002] The current deep neural network is developing rapidly and its application is becoming more and more extensive. The method of using deep neural network for target detection or search on video or image mainly includes the two-step method represented by Faster R-CNN, R-CNN, etc. and YOLO , SSD, etc.; although Faster R-CNN is an excellent algorithm in the two-step method, it can only achieve a processing speed of 5FPS with the support of the powerful computing power of the K40 GPU, which cannot meet the real-time requirements; although the one-step method The speed of YOLO and SSD target detection can reach more than 15FPS to meet the real-time requirements, but it must be supported by the computing power of Titan X or M40GPU. The algorithm with better pe...

AI Technical Summary

Problems solved by technology

[0003] UAV is a reusable aircraft controlled by radio remote control or autonomous program, unmanned, it has simple structure, low cost, strong survivability, good maneuverability and can complete various types of tasks Advantages; but the low weight of the UAV prevents it from carrying powerful computing equipment, which makes it difficult to deploy target detection algorithms based on deep neural networks, and the weight of small UAV onboard computers such as Raspberry Pi or odroid Lightweight and limited computing power; even if TinyYOLO or Mobilenets-SSD in the faster one-step method is deployed on the odroid airborne computer, the target detection speed does not exceed 3FPS, which cannot meet the real-time requirements

The decommissioned Predator UAV mainly uses the UAV's sensor to obtain data and send it back to the ground for manual interpretation on the ground; the improved "Global Hawk" can carry signal sensors and radars for detecting ground moving targets. On-board target detection and monitoring capabilities (distinguish between motion and static, detect moving targets), and the detection technology is not mature enough; Rainbow drones obtain data from the drone's sensors and send them back to the ground, where they are manually interpreted on the ground and further processed at the back end; artificial intelligence algorithms are used in " Scanning Eagle” was tested. Only a few days after the start of the test, the accuracy of the computer’s recognition of people, vehicles, buildings and other objects reached 60%, and it increased to 80% after 1 week. However, this application is still done on the ground From this point of view, the current technology is still unable to track and detect the target in the data collected by the drone's onboard camera in real time and perform the processing operation of the next step instruction

Images