Algorithm Design And System Implementation Towards Real-Time Object Detection On Embedded Platforms

As one of the basic problems in the field of machine vision,object detection has received extensive attention in the past decades.With the recent advancements of embedded technology and embedded devices,object detection on edge platforms has become a research hotspot,and has been widely adopted in various fields and tasks.Currently,deep neural networks（DNNs）based end-to-end detection models have achieved far better performance than traditional methods.However,these models feature with high complexity and large amount of computation,and cannot meet the requirements of low power and real-time processing.Thus,they cannot be directly deployed on embedded devices.To tackle the aforementioned problems,we present an energy-efficient real-time DNN-based detection system for embedded GPU platforms.We focus on the scenario of Unmanned Aerial Vehicle（UAV）based single object detection,and reach the solution from data analysis,algorithm design to system implementation.To be specific,a basic detection framework is first designed based on the single target detection task.The detection framework consists of four sub-modules,and supports two modes of training and prediction.Then,according to the characteristics of the UAV-based single object detection task,the basic detection framework is deeply optimized from the aspects of detection scale selection,network architecture optimization,training loss optimization,and so on.After algorithm-level design and optimization,an detection system is implemented on the Nvidia Jetson TX2 embedded GPU platform based on the well-trained detection model.Finally,the detection system is further optimized with several parallel technologies to achieve real-time processing.A large-scale UAV detection dataset（LPODC）is used for training and performance evaluation for detection system.In the local evaluation,the proposed system can achieve accuracy of 0.886 Io U,with the speed of 28.46 FPS and the power consumption of 10.640 W at Max-P-Core-All mode on Nvidia Jetson TX2 platform.In the official evaluation of the 2018 DAC System Design Contest,the proposed system can achieve accuracy of 0.691,with the speed of 25.30 FPS and the power consumption of 13.27 W at Max-N mode on Nvidia Jetson TX2 platform,which also won the 2^nd place among 53 systems in the comprehensive evaluation of accuracy,power consumption and speed.