Research On Targrt Detection And Tracking Of Monocular Vision Image For Gyroplane

The application scenario of common fixed monitoring equipment is single,and the deployment cost is high in the detection of traffic conditions.Gyroplane has the advantages of small size and high flexibility,which can carry out image processing tasks in the air,and the monitoring of various scenarios is no longer limited to the fixed platform.However,there are still some shortcomings in the detection and tracking algorithms applied to the rotorcraft platform at this stage.First,the current detection and tracking algorithms are mainly based on deep learning,and the network structure is complex.Due to the limited performance of the airborne equipment suitable for the rotorcraft platform,the requirements of computing resources and other related requirements cannot be satisfied.And most of the small targets in the perspective of rotorcraft aerial photography lead to the lack of image feature information.It is difficult to extract information,which causes problems such as missing or wrong detection or brings great challenges to the detection algorithm.To solve the related problems,the main research work has the following three points:(1)Based on the YOLO series detection algorithm,the lightweight improvement is carried out.Aiming at the disadvantage of low computing power of gyroplane platform,the network structure is optimized.The sparse channel pruning method based on BN layer is used to lighten the network of YOLO-V3 which has high performance but large model.At the same time,the prior frame is redefined by K-means++ algorithm to improve the detection efficiency of aerial target.The improved network accelerates the reasoning speed,reduces the volume of the model,and improves the detection accuracy of aerial targets.The low performance small network model YOLO-tiny is enhanced to increase the size of convolution layers of the network and suppress the increase of the model by using the NIN layer.The detection accuracy is improved on the basis of maintaining realtime performance.(2)Deep SORT tracking algorithm is combined with the improved YOLO detection algorithm.Aiming at the situation of missing detection,combined with the characteristic information of the target in the time series in the detection results,the results of the improved YOLO detection algorithm are used as the input of the deep sort algorithm tracker,and the position of the target is predicted by using the relationship between the front and back frames.At the same time,it can increase the accuracy of detection and reduce the error detection rate.(3)A target detection and tracking system based on monocular vision is built on the platform of M100 gyroplane.The DOTA-like data set is made,and the improved algorithm is deployed on the airborne and ground experimental platforms respectively.Finally,the flight test is carried out at ultra-low altitude,and the multi-target detection and tracking tests are carried out under different degrees of background interference.The experimental results show that the system has good stability and reliability in the analysis of pedestrian and vehicle traffic.