| Target detection is one of the key technologies in intelligent security,automatic driving,defect detection and other popular fields.With the development of deep learning,target detection technology based on deep learning has achieved excellent results and has a good theoretical foundation.Target detection have a wide range of requirements for embedded devices,and the embedded GPU platform has become one of the most commonly used image processing platforms because of its excellent computing and the convenience of model deployment.However,the current target detection based on deep learning has a high demand for embedded devices and can not be fully applicable to the embedded GPU platform.Therefore,this Thesis studies the implementation of object detection algorithm on embedded GPU devices.The main work and innovation of this Thesis are as follows.On the basis of combing the existing target detection algorithms,this Thesis selects efficientdet as the baseline model and makes lightweight improvements to make the detection algorithm suitable for embedded GPU devices with different resources.Then,the efficientdet is lightweight.Firstly,the baseline model is improved to the method without anchor box,and the scheme of multiple positive samples is used to alleviate the problem of sample imbalance.Then the target probability prediction graph is added in the detection head by using the method of weight sharing,which Speed up post-processing and improve detection accuracy.Finally,the loss function of the baseline network is optimized to make up for the loss caused by lightweight.The compression and acceleration methods of the model are studied,and its deployment is completed.The compression model uses two methods: pruning and quantization,which greatly reduces the number of parameters on the basis of ensuring that the map decline is within an acceptable range.Then,through the horizontal and vertical combination of the model structure,the relationship between bandwidth and GPU processing capacity is reasonably used to speed up the prediction speed.After the experimental verification of coco dataset,the result parameters are only21.19% of the baseline model,and only 2.06 map values are lost.Good experimental results are obtained.The experimental data on Jeston Xavier NX platform shows that the detection speed of the lightweight model is 5.32 times higher than that of the baseline model.Finally,based on the above,an object detection application system based on embedded GPU platform is developed.The system uses the Jeston Xavier NX embedded GPU platform.Users can conveniently detect outdoor signboards and save the corresponding detection results through the system,which can complete the task of target detection while ensuring the accuracy and detection speed. |
