Design Of Portable Object Detection System Based On Deep Learning

Object detection is mainly to determine whether there is an object of interest in the image,and its precise positioning.It can be applied to object detection,number statistics,object tracking and so on.The traditional object detection algorithm generally uses slidin g window to select the image region,and using SIFT,HOG or other methods for feature extraction,and using SVM,Adaboost for category judgment,there are poor robustness of feature extraction,regional selection without targeted,high time complexity,and so o n.The object detection model based on deep learning,such as SSD,faster R-CNN and YOLOv3,has been greatly improved in terms of robustness,speed and precision,and the main problems are:the model training needs powerful computing power,the model depl oyment is cumbersome,and the model inferencing still needs some computing power.In this paper,a portable object detection system was studied,and a high performanc e and low power neural computing accelerator was added to the embedded system with 1 imited computing power and power,and the main deep learning model inference was car ried out to realize fast object detection.The main work includes:(1)The overall design of the portable object detection system was determined throug h investigation.YOLOv3 model was selected as object detection algorithm,and Intel NC S2 was selected as inferencing accelerator,and OpenVINO was selected as model conver sion,optimization and deployment,and RaspberryPi 3b+was selected as embedded platfo rm?(2)YOLOv3 based object detection prototype system was built on PC,and the syste m performance was evaluated.The system uses Intel E5-2670 CPU+NCS2+USB Ca mera+Ubuntu16.04+OpenVINO+YOLOv3/tiny-YOLOv3 structure.With a single NC S2,YOLOv3/tiny-YOLOv3 inferencing speed is 1.6/14.3 fps,it's about 43%inference pe rformance comparing to E5-2670 CPU.(3)The portable object detection system platform was built with RaspberryPi 3b+(ARMv7 CPU)+NCS2+USB Camera+Raspbian stretch+OpenVINO+YOLOv3/tiny-YOLOv3 structure.With a single NCS2,YOLOv3/tiny-YOLOv3 inferencing speed is 1.2/7.1 fps.Y OLOv3/tiny-YOLOv3 model increases system power consumption 1.4w/1.2w regardless of the platform itself and external touch screen,and it's suitable for portable scenarios.The model can be retrained for the specific application scenario.The result showsthat the perf ormance of the model is similar with the same network structure and different weights.