Design And Implementation Of Pedestrian Detection And Tracking System From UAV Based On Lightweight Convolutional Neural Network

In recent years,UAV has been widely used in various fields such as transportation,emergency rescue,agricultural production and military operations,and has shown a trend towards intelligence,autonomy and high reliability.The object detection is vital as the"eyes" of the UAV With the flourishing development of artificial intelligence technology,neural network methods in object detection have become the preference with high robustness and low time complexity,compared to traditional object detection methods such as HOG and SIFT,have avoided sliding window redundancy and manual feature extraction for higher detection accuracy,stronger robustness and better feature representation.However,the memory and computational power of airborne embedded devices is limited,and common CNN models are usually difficult to deploy and apply in engineering applications.Therefore,an object detection and tracking algorithm based on lightweight conventional neural network was designed for embedded platforms with limited computing power.It is of great significance for the expansion of deep learning machine vision algorithms on UAV platforms.The following research is carried out:(1)This thesis designs and implements an improved object detection algorithm based on YOLOv4-tiny.Comparing and analysing multiple deep learning-based object detection algorithms,using the speed-first principle,the less computationally intensive YOLOv4-tiny was chosen as the base model.The model parameters are filtered and censored using network pruning,and Auxiliary_Block is added to compensate for the detection accuracy.The models are trained and tested using public datasets and specific experimental environment datasets to verify effectiveness.(2)This thesis designs and implements an improved target tracking algorithm based on Deepsort.A feature extraction network based on ShuffleNet v2-0.5 is proposed for the original network which has too many parameters and runs too slowly on the ARM platform.It is jointly trained based on multiple ReID datasets to obtain the best weight parameters.The thesis compares the metrics of the before and after model on TX2 to verify that the method has significantly improved the performance of the deepsort target tracking algorithm on the ARM platform.(3)This thesis build a UAV ground object detection and tracking system.Considering the UAV load,power consumption and other factors,this paper designs and implements the hardware platform and software process of the UAV ground object detection and tracking system.Finally,the thesis uses public and real datasets to detect and analyse the system online.The UAV pedestrian detection and tracking system designed and implemented in this project can run successfully on both PC and embedded platforms.The experimental results show that the pruning and optimization of the convolutional neural network structure can significantly reduce the computational parameters,increase the system operation speed,reduce the memory usage,and take into account the detection and tracking accuracy.The results of this research provide ideas for the deployment of convolutional neural networks on low-power embedded platforms and have high engineering value.