Research On Object Detection Method Of Embedded System For UAV

Object detection is a key task in the field of computer vision.The object detection of UAV oriented embedded system is an important scene in the application of computer vision technology in practical engineering.The problem of small object and embedded device deployment in the scene poses a great challenge in practical application.Therefore,how to design a model that can accurately detect small object in UAV dataset and facilitate deployment in embedded devices with weak computing capacity and tight memory space has become a hot topic.In order to solve the object detection task of the embedded system for UAV,this paper proposes a lightweight model for small object detection,and compresses the model,so that it can have a good detection effect on UAV dataset,and can be well deployed on the embedded system.The details are as follows:To solve the problem of large amount of computation and weak learning ability of backbone network,this paper proposes a cross stage partial residual module.The residual structure can solve the problem of gradient disappearance,and the cross stage partial connection structure can reduce the network computation and improve the network learning ability.Taking this module as the backbone network can speed up the network reasoning speed and improve the detection accuracy.Aiming at the problem of poor detection effect of small object,this paper proposes a multi-scale path aggregation module.The module can predict the feature map of different scales,and can integrate the semantic information of deep feature map into the shallow feature map and the location information of the shallow feature map into the deep feature map.Make full use of the detailed information of the feature map of each layer of the neural network,enrich the multi-scale features of the network,and improve the detection accuracy of small object.In order to solve the problems of weak computing power and tight disk and memory space of embedded system,this paper proposes a model compression method based on channel pruning and knowledge distillation.Channel pruning can eliminate the channel with low weight and little influence on the accuracy of the model,effectively reduce the parameters of the model,and then compress the volume of the model.Knowledge distillation,as a strategy to guide training,takes the model before pruning as the teacher network to guide the student network after pruning,which can have a good effect.