Research And Implementation Of Model Compression For Object Detection And Segmentation In Road Scenes

With the development of science and technology,especially the acceleration of hardware platforms,deep learning has achieved remarkable results in many fields.Among them,autonomous driving has become one of the most popular research directions.Autonomous driving is a complex subject that integrates multiple research disciplines,including object detection,semantic segmentation,instance segmentation,object tracking,reinforcement learning,and more.Among these technologies,object detection and semantic segmentation play a more important role as the basis of other tasks,and their performance has a great impact on the entire autonomous driving technology.However,the current target detection and semantic segmentation network models are basically faced with problems such as a large amount of parameters,high computing power requirements,and long prediction time,which makes these network models do not get an expected performance on devices with limited computing power,such as embedded devices.In this thesis,the general scene of automatic driving--road scene is taken as the research basis,and the model compression on the algorithm models YOLOv4 and YOLOv5 that get a better performance in target detection at this stage and the algorithm model Deep Labv3 p which gets a better performance in semantic segmentation are performed on this basis.On the premise of ensuring the prediction accuracy of model,the model compression algorithm reduces model’s parameter size and inference time so that the model can achieve a good performance on embedded devices with limited computing power.The main work of this thesis is as follows:（1）The commonly used road scene datasets are analyzed,and the detection model and segmentation model are optimized according to the analysis results.Optimize the YOLO model by adding Asymmetric Loss to its loss function and improving its anchor frame structure,which improves its m AP on the Cityscapes and BDD100 k datasets.Optimize the Deep Labv3 p model by adding improved RMI Loss to its loss function,which improves its m IOU on the Cityscapes dataset.（2）Combined with a variety of model compression methods to compress the target detection models—YOLOv4 and YOLOv5.The YOLOv4 model reduces the number of parameters by about 69% and increases prediction speed by 46 fps with only a loss of 2.5%of m AP.The YOLOv5 model reduces the number of parameters by about 70% and increases prediction speed by 37 fps with only a loss of 1.3% of m AP.Combined with a variety of model compression methods to compress the semantic segmentation model—Deep Labv3 p.The Deep Labv3 p model reduces the number of parameters by 22 times by replacing its backbone network.The prediction time of the small model after replacing the backbone network is reduced by about 50% and the its m IOU is increased by 2% by combining quantification and knowledge distillation.（3）Based on the Pytorch deep learning framework,a complete model compression process code is written,including initial training,model compression,and result analysis.Combined with Py Qt5,a model compression system for a target detection model and a semantic segmentation model based on road scenes is designed and built,which increases the ease of use of these algorithms.