Accurate and efficient traffic participant detection based on optimized features and multi-scale localization confidence

Abstract

Traffic participant detection, which locates targets participating in road traffic such as vehicles and humans, is a critical research goal in autonomous driving and intelligent transportation technology, in which targets of various scales are challenging to handle. To address this problem, we propose an accurate and efficient traffic participant detector called FM-RepPoints to increase detection precision of multi-scale targets via optimized features and localization confidence. We first present a parameter-efficient backbone Dilated-ResNeSt to extract complicated urban scene features with less inference time. An adaptive attention module is then added to calibrate the features according to the corresponding targets. Moreover, we set the multi-scale localization confidence for traffic participants of various scales based on the sensitivity analysis. The proposed network outperforms the basic network by 3.3 points mAP; the accuracy of vehicles and humans is improved by 4.5% and 2.1%, respectively. Compared with state-of-the-art algorithms, the proposed method achieves the highest accuracy while maintaining average inference speed, thereby accomplishing the best speed-accuracy trade-off.