Speed Accuracy Trade-off in Pedestrian and Vehicle Detection Using Localized Big Data

Abstract

This paper aims to obtain a guide of several experimental object detection architectures for pedestrian detection / vehicle detection (PD/VD) in the South Korean driving environment and to present the trade-off relationship between the mean average precision (mAP) and the frame per second (FPS) values. For these purposes, we generated a Korean vehicle black box front view dataset (KVD) to consider the actual driving environments in South Korea. We then experimented with various configurations of object detection architectures using the KVD. Next, the trade-off relationship for each architecture was summarized and analyzed. This paper presents a guide for choosing PD/VD architectures to achieve a suitable mAP–FPS balance for advanced driver assistance systems (ADAS) and autonomous navigation technologies.