Research And Implementation Of Real-time Object Detection Algorithm Based On Deep Learning

With the development of Internet of Things technology,the amount of digital images and videos has increased explosively,and image data analysis is facing more and more challenges.Object detection is an important image processing technology,its main task is to classify and locate the object of interest in a picture.At present,object detection technology based on deep learning does not achieve a good balance in precision and speed,and the huge amount of computation makes it difficult to achieve real-time inference speed on embedded devices.This thesis focuses on the research of real-time object detection algorithm,and explores the inference optimization scheme on embedded devices.There are three main contributions.First,a real-time object detection algorithm called ARFSSD(Attentional Receptive Field Fusion Single Shot MultiBox Detector)which based on visual attention mechanism is proposed.ARFSSD integrates RF-DRN(Receptive Field Based Dilated Residual Networks)and CAM(channel attention module).RF-DRN systematically aggregates multiscale contextual information by increasing or fusing multiscale receptive fields.CAM reduces the correlation between feature channels by redistributing channels' attention.Second,a lightweight object detection network ARFSSD-Lite is designed for NVIDIA Jetson TX2 embedded device,which reduces the computational load by reducing the depth of feature extraction network and the resolution of input images.Third,by using CUDA and TensorRT,a real-time object detection algorithm inference optimization scheme is proposed.A real-time object detection system is deployed on Jetson TX2 through CUDA multithreading,storage management,communication optimization,quantization and custom plugin layer.Experiments show that both the real-time object detection algorithm and the inference optimization scheme have achieved the expected goals.For VOC2007 datasets,ARFSSD achieves 78.6% mAP,and running at 109.9 FPS,which shows that the precision and speed have been improved.In addition,ARFSSD-Lite achieves 73.7% mAP and 217.8 FPS on VOC2007.For Jetson TX2,after inference optimization,ARFSSD-Lite only lost 0.6% mAP on VOC2007 and the speed was increased from 16.7 FPS to 35.2 FPS.More importantly,ARFSSD-Lite achieves 90.6% mAP and 47.6 FPS on self-built bread datasets,which further proves the effectiveness of the algorithm and inference optimization scheme.