http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
임베디드 시스템 기반 실시간 물체 검출을 위한 YOLOv3 최적화
위성민(Seong Min Wi),최경택(Kyoungtaek Choi),정호기(Ho Gi Jung),서재규(Jae Kyu Suhr),김도윤(Doyoon Kim) 한국자동차공학회 2021 한국자동차공학회 부문종합 학술대회 Vol.2021 No.6
A local dynamic map(LDM) is one of the key components for autonomous driving. Surveillance cameras are frequently used to detect moving objects and transfer this information to the LDM. To this work, it is mandatory to run an object detection algorithm in a real-time embedded system. This paper introduces a way to optimize a deep neural network(DNN)-based object detector in terms of computational cost, memory space, and inference time. This paper uses two techniques for optimization. One is network slimming that can prune less important filters from the network. The other is a quantization-aware training that can change weights of the network from float32 to int8. With the help of these two techniques, a well-known object detector, YOLOv3, has been optimized and run in real-time using the DSP of Qualcomm QCS605.