Research On Model Acceleration Techniques For Aerial Object Detection

Unmanned Aerial Vehicle(UAVs)are widely used in intelligent transportation,agricultural production,rescue,and disaster relief due to their high mobility and ease of use.With the rapid development of deep learning,UAVs equipped with cameras to complete computer vision tasks have become one of the mainstream research directions.Object detection in aerial images scenes can not only complete the practical tasks,but also serve as the basis for other computer vision tasks on the UAV platform.However,the edge device platforms carried by UAVs often have been limited in memory,power consumption and arithmetic power to meet the deep learning model deployment requirements.Therefore,the research on deep learning-based model acceleration methods in aerial scenarios is important for expanding the application of object detection algorithms in mobile devices.In this thesis,aiming at the application of object detection algorithm in aerial photography scenes,the model pruning and structural re-parameterization techniques were used for model acceleration,And the further acceleration by Tensor RT deployment.The main work and contributions of this thesis are summarized as follows:(1)For the limited and difficult access of open infrared aerial photography dataset,a selfbuilt infrared aerial photography pedestrian dataset Air-pedestrian is constructed and the associated data distribution is analyzed.The YOLO object detection algorithm model is trained and tested on this dataset in preparation for the subsequent model pruning work.(2)To obtain a better speed-accuracy balanced model,a compact structure IRA-YOLOv5 for aerial object detections is proposed.Embedded platforms are liming computing resources for deployment,and most of the current deep learning algorithms are parameter redundancy.In this thesis,based on the constructed YOLOv5 algorithm,a sparse model is constructed by introducing Smooth-L1 regularization,and channel pruning is carried out by identifying the importance of feature channels by scale factors to reduce the amount of training parameters and computation of the algorithm.The experimental results of model evaluation based on the self-built aerial infrared dataset Air-pedestrian show that the mean average accuracy(m AP)of IRA-YOLOv5 is 96.4%,which is comparable to the original YOLOv5 model.The model volume is compressed by 95.8%,the number of parameters is reduced by 98.3%,and the inference acceleration on the PC side is about 30%.The onboard Jetson TX2 inference frame rate has been increased from 14.5 FPS(Frame Per Second)to 19.2 FPS.Compared with YOLOv5,the proposed model achieves the better balance of detection speed and accuracy.(3)Based on YOLOv5 algorithm and structure reparameterization optimization,an object detection model Rep-YOLO for aerial photography scene is proposed.During training stage,the Rep VGG module with multi-branched structure is introduced in the backbone network to improve the feature extraction ability,and the multi-branch structure of the Rep VGG module is re-parameterized during inference to reduce the network branch and structural complexity.At the same time,combined with the data prior characteristics,the path aggregation network at the neck of the detection network is optimized to improve the precision-speed balancing ability.Comparative experiments were performed on two public available infrared datasets.The statistical results Com Net dataset show that the proposed Rep-YOLO algorithm improves the detection accuracy,while the parameters and model size are reduced by 29.7%and 23.2%,respectively.It provides reliable technical support for the improvement of detection algorithm and practical application of aerial photography scene.(4)To further achieving model acceleration,the Tensor RT framework is used to deploy algorithm models on the airborne Jetson TX2 embedded platform.Experimental results show that the deployment of deep learning models on the Tensor RT framework can significantly reducing model inference time and graphic memory with comparable inference accuracy,which meets the application requirements of deep learning algorithms in UAV platforms.