Research On Object Detection Optimization Algorithm Based On Deep Learning And Application On Embedded Computing Platform

In recent ten years,deep learning has achieved remarkable results in both theory and engineering,such as image recognition,target detection and natural language processing.However,the feature extraction part of the traditional target detection needs artificial design of features before deep learning is applied to target detection The features of artificial design of features become very difficult when we face diverse and complex scenes of target features.Target detection based on deep learning does not need to design artificial features,it just use convolutional neural network to learn.From R-CNN,SPP-Net,Fast R-CNN,Faster R-CNN,R-FCN based on regional nomination to YOLO and SSD which are end-to-end,the recognition accuracy and rate have far exceeded the traditional methods.It has now been used in driverless driving,vehicle detection pedestrian detection and other fields.At present,most target detection models based on deep learning are too large to realize forward reasoning directly on embedded computing platform.And the common working mode is "embedded computing platform--cloud computing--embedded computing platform".It means the images or videoes which need to be detected is collected on the embedded computing platform.Then,it is transmitted to the cloud server through the network and detected by the target detection model based on deep learning.The detection results are finally transmitted back to the embedded computing platform,which will not only lead to excessive dependence on the network,but also cause the delay of the result displayTo solve this problem,this paper studies the target detection optimization algorithm based on deep learning.The main research contents of this paper are construction of lightweight network,application of model pruning and optimization of model assembly.Firstly,an appropriate lightweight neural network and target detection algorithm are selected to build the network,then to train a small size and high precision target detection model with own data.Then,the trained model to be pruned.While shortening the time of forward reasoning,keep the accuracy constant as far as possible.Finally,we transplant the target detection model which has been pruned to the embedded computing platform and then use convolution algorithm and NEON technology to realize forward reasoning acceleration.This paper designs and implements a target detection reasoning system based on Raspberry PI,an embedded computing platform.Firstly,the embedded computing platform--Raspberry PI was used to collect data,and then data annotation,model training and model pruning were completed on the PC.Finally,the model was transplanted to the embedded computing platform--Raspberry PI,and the performance of target detection and inference system was improved through assembly optimization.Finally,it is proved that the system achieves the expected target of target detection directly on the embedded computing platform through the test.