Research About Lightweight Object Detection For Edge Computing

In recent years,object detection has been widely used in many fields such as Intelligent Safety System,Intelligent City,Intelligent Plant,Intelligent Vehicle,and so on.Since the core of object detection is deep learning,at present object detection models are mostly deployed on server side,which have two shortcomings: first,with the increasing demand in deployment,the way to configure a lot of high-performance servers is facing the enormous economic pressure;second,the data collected by the edge device such as cameras will be transmitted to server through internet.If the data is large enough,it will cause the detection delay.With edge devices becoming more and more popular,advantages such as small size,high cost performance,and low power consumption make edge device becoming a new choice for object detection operating environment.Like Faster RCNN,SSD,and YOLO,excellent object detection models can achieve high precision.Because of its high requirements of calculations,the detection speed in the operating environment with poor performance such as edge devices can't fulfill the demand of industrial applications.Model optimization techniques have maturely developed,such as transfer learning,model pruning,and fine-tuning,which could be of help to optimize model and improve model performance on edge devices.Under such circumstances,we proposed a general method on optimizing lightweight object detection model for edge calculation environment.We took a specific application as an example to verify the effectiveness of this method.Detailed work is listed in the following.(1)Presented a compact software environment,which is suitable for object detection on edge devices.Considering the limited resource of edge devices,we picked out the right one from many deep learning frameworks and object detection models,and constructed a compact software environment for edge devices.(2)Designed a neural network pruning method for general object detection,based on transfer learning.First,we analyzed the object classes which could be detected in actual scenario,then we cut down the unrelated classes among the general object detection model by transfer learning.This method can slim the model in the horizontal direction,and significantly speed up the detection without precision decreasing.(3)Designed a method to pruning layers,according to object scale which would be detected.Based on the detecting object scale in actual application,we investigated layers that detect different scale objects,and cut down the layers which detect too small or too big scale objects.This method can shorten the model in the vertical direction,and improve the detection speed.(4)For the object detection model pruned above,we optimized the data set accordingly.There was too little data matching the actual application scenario among the standard dataset,then we increased the data by two ways,to help the model converging quickly during training.The first one was to choose more image data from the standard data set that matching the actual application scenario,then increase data further by data augmentation.The second one was to create customized data manually,which means collecting the data directly in the actual application scenario,and also increase data further by using data enhancement method.In this paper,NVIDIA Jetson TX2 board(8GB memory)was chosen as the representative of edge device.SSD300(VGG16)was selected as the lightweight object detection model from three aspects: model size,detection precision and detection speed.First,the high-precision model such as Faster RCNN couldn't run successfully on the edge device,as it would be out of memory.Second,SSD(VGG16)was better than other lightweight object detection model,such as SSD(Mobilenet),considering the tradeoff between precision and speed.By transfer learning and the layer pruning method,we cut SSD to detect the "person" class only,and then cut down the conv7₂,conv8₂ and conv9₂ layers which detecting large-scale objects.Optimized model reached a speed of 2.42FPS(0.413sec/frame),comparing to the original model with 1.56FPE(0.64sec/frame),the 55% acceleration was achived.We trained SSD with the optimized standard dataset MSCOCO2014-Human and the custom-made dataset collected in the actual scenario.The detection precision reached 90%,which acquired an increasement of 20%,compared with original precision 70%.The experimental results showed that the general method to optimize lightweight object detection model for edge environment proposed in this paper has significant effects.