Target Recognition And Visual Positioning Techniques Based On Deep Learning

Unmanned aerial vehicle(UAV)for ground target identification and visual positioning technology,the backlog of target recognition accuracy and speed directly affects the subsequent target,the effect of using unmanned aerial vehicle(UAV)visual sensor collected most of the ground target vehicle images belongs to the category of small targets,and in the field of vision,the early development of small target detection technology from the field is still a worthy of research to overcome difficulties.The traditional edge detection algorithm has low accuracy in processing the aerial image of small ground target.The traditional feature operator detection algorithm is unable to realize the real-time performance of UAV visual positioning technology due to the fact that too many feature points need to be extracted in the block-matching strategy and the matching time becomes longer.Aiming at the problems of low accuracy of target detection,difficult detection of small targets and large positioning error in the traditional UAV ground target recognition and visual positioning technology,this paper studies the target recognition and visual positioning technology based on deep learning.Put forward a kind of based on improved DSOD(Deeply Supervised Object Detectors)UAV visual positioning method,Solve the problems of complex background and poor detection performance of multi-angle small target in UAV aerial target image.By building DSOD network model,this technology can realize the training model of UAV aerial photography ground vehicle target detection network from scratch,generate the globally optimal training model on detection task,and reduce the complexity of network structure and the number of parameters.Dense Net(Dense-Connection Net)is used as the trunk feature extraction network to realize the intensive connection of the front and back layers of the network,which makes the feature information of the output layer richer by virtue of feature reuse.Deconvolution and forecast module design expanded testing scale,design of modulation coefficient optimization of training effect,merge network speed up these points up to faster small target detection vehicle identification precision and real-time visual positioning,will get the goal of bounding box pixel coordinates calculated through coordinate transformation finally pending the position information of ground targets.The experimental results show that the improved DSOD target detection network can be used to improve the detection accuracy of small targets on the ground vehicle,and the average error of target positioning on the ground vehicle can be controlled within0.29 m in the range of 80-120 m flight height,and meanwhile,the real-time performance of UAV visual positioning can be improved.