PBTA: Partial Break Triplet Attention Model for Small Pedestrian Detection Based on Vehicle Camera Sensors

Successfully detecting small pedestrians through vehicle camera sensors would greatly facilitate the development of autonomous driving safety applications. However, the existing pedestrian detection models applied to vehicle camera applications were limited by the scale confusion problem and the weak feature problem of small pedestrian targets. To resolve these issues, this study proposed a partial break triplet attention (PBTA) network composed of two components: the partial break bidirectional feature pyramid network (PBFPN) and the TR-NCSPDarknet53. PBFPN was used to solve the scale confusion problem in shallow feature maps by employing a partial break operation and a branch fusion operation. In TR-NCSPDarknet53, the Ta-conv module was proposed to solve the weak feature problem. The PBTA network provided a new improvement idea for small pedestrian detection. It greatly improved the accuracy while keeping the parameters at a low level, which is vital for safety applications in autonomous driving. Extensive experiments on CityPersons, Crowdhuman, and WiderPerson datasets including various traffic images from camera sensors demonstrate the accuracy of the PBTA network in small pedestrian detection. Compared with the baseline (Yolov
$8_{\text {S}}$
) network, the accuracy of small objects (APS) is improved by 50%.