Design And Implementation Of Object Detection System On Mobile Device

Object detection is a basic research field in computer vision,which can help computers understand what the images or videos mean by detecting and recogniz-ing objects in them.With the development of deep learning,this technique has made a big progress these years.However,almost all the deep learning based ob-ject detection algorithms are designed to be used in servers,which occupy a large amount of memory and the speed is too slow.While on embedded devices,the computing resources are limited,it is hard to run such a complexed algorithm.This paper designs and implements a completed and light-weighted object detect framework based on mobile devices,which mainly focus on neural network model compression and fast computation.Single Shot Detection(SSD),one of the most polular object detection algorithms currently,is chosen to be the base algorithm and many modifications are made to minimize the model size and speed up the inference time.Firstly,we design a light-weighted CNN inference framework in order to deploy a detection model trained on server into embedded devices.Especially,on iOS system,Neon instruction optimization and mobile GPU are used to accelerate the convolution computation.For model size,a 5-bit quantized Tiny-Darknet backbone network is used to replace the original VGG16 network,which makes the model size reduced over 70x.For accuracy,we use multi-scale feature aggregation methods such as dilated-convolution and reverse connection.Finally,Staple object tracking algorithm is used to help improve the fluency and stability.We develop a demo application based on iPhone7 Plus,which can capture real-time video from camera,detect and track the objects and draw them on the screen.This application can detect and track at most 5 objects at the same time.The model size is 1.4MB,and the whole average speed of the system is over 35FPS.