Research Of Indoor Multi-Moving Target Tracking Technology Based On Deep Learning

Compared with traditional target detection technology,deep learning has excellent detection accuracy and generalization ability.Its application in multi-rotor UAV can significantly improve the intelligence of UAV,realize the recognition and tracking of multiple complex targets in large scene and complex situation,and has good application prospects.In this paper,deep learning algorithm is applied to multi-rotor UAV to realize tracking of multiply moving targets in complex indoor environment.The hardware platform is based on NVIDIA TX2 core board,and an airborne deep learning computing platform is designed.The target detection algorithm is based on YOLO algorithm,aiming at the UAV applications,the computing resources,detection accuracy and detection speed are optimized.First of all,this paper analyses the task requirements of the subject,and designs a target detection and tracking scheme based on deep learning algorithm.In aspect of hardware,a small volume and lightweight airborne deep learning computing platform is designed based on NVIDIA TX2 core board.In aspect of algorithm,the principles and performance of several prevalent deep learning target detection algorithms are compared and analyzed.Then,a deep learning detection algorithm for specific ground targets is studied and developed based on YOLO algorithm,due to the balance between detection accuracy,speed and resource consumption.In this paper,one data set of ground targets is established,the network parameters are adjusted and trained,and tested on the data set.The test results show that the algorithm developed in this paper has better performance for ground targets in complex environments than the traditional machine vision algorithm.In order to improve the performance of deep learning target detection algorithm in multi-rotor UAV,the optimization methods are studied from two directions of detection accuracy and detection speed.In terms of accuracy,this paper uses multi-scale prediction mechanism and K-means-based target bounding boxes clustering algorithm to optimize the detection performance for small area targets and improve the recognition rate and location detection accuracy.In terms of speed,this paper uses TensorRT accelerated inference and INT8 weight quantization to optimize the deep learning algorithm.Results show that the above measures can improve the detection speed more than twice.Finally,this paper studies the indoor multi-moving target tracking technology.Firstly,for the target position detected by the deep learning algorithm,the continuous and stable target position is obtained by the particle filter algorithm.Then,the transformation relationship between the airborne camera and the global coordinate system is analyzed,and the indoor positioning technology of UAV based on visual information is designed.Combining the two methods,the global coordinates of the vehicle itself and all moving targets are set up,and the experimental simulation environment is built to verify the effectiveness of the proposed target detection and tracking algorithm.