Design Of Real-Time Object Detection System Based On Embedded Heterogeneous GPU Platform

Object detection is widely used in the field of information technology and artificial intelligence,including robot vision,autonomous driving,intelligent monitoring on unmanned aerial vehicles(UAV)and so on.A variety of application scenarios put forward an urgent demand for real-time object detection with high precision and high energy efficiency on embedded devices.Object detection is one of the most challenging problems in the field of computer vision,it not only needs to classify objects in the image,but also locate them.Traditional solutions usually adopt the sliding window approach and use the trained classifier to judge all the possible windows,which leads to significant limitations in terms of not only high computational complexity but also high detection error rate.With the continuous development of deep learning,the realization convolutional neural netvork(CNN)-based methods achieves increasingly high performance.Particularly,some end-to-end object detection models such as SSD and YOLO have made a new breakthrough in the speed of object detection.Although the current object detection technology has a very high detection accuracy and speed,when applied to embedded platforms with limited computing and memory resources,the performance of these technologies is greatly reduced and can not meet the real-time standard.Therefore,it is still a great challenge to realize real-time object detection on embedded platform.This paper chooses NVIDIA JETSON TX2,an embedded heterogeneous GPU platform,to carry out the research.On the basis of learning and summarizing the technology of object detection,several models with higher performance are deployed to TX2 for comparative analysis.After weighing the two important indicators of accuracy and speed,the YOLOv2 model is chosen to optimize and research,so as to achieve the real-time object detection system on embedded platform.The optimization scheme mainly includes the following aspects:First,modify the YOLOv2 network structure and parameters to accelerate the network.Second,adopt half-precision floating-point numbers to reduce the amount of computation in the network.With this method,the speed is significantly improved at a slightly cost of the accuracy.Third,analyze the time consumption on the CPU and GPU during the detection and design the multi-threaded pipeline that obtains the speed improvement of 4FPS.The above three optimization schemes are mainly designed and implemented according to the structure of the network itself and the characteristics of the actual platform,and are applicable to different data sets.In addition,according to the requirement of real-time detection of moving objects to achieve tracking,we combine optimized YOLOv2 with the tracking algorithm.It is found through experiments that that when detecting moving objects in video,the combination of GOTURN target tracking algorithm and optimized YOLOv2 makes the system achieve higher system performance.