汽车智能座舱遗留物品检测

Leftovers Detection in Auto Smart Cockpit

随着汽车产业数字化、智能化发展,智能座舱已成为未来的趋势。本文针对车内遗留物品检测任务开展研究,分析车载场景图像数据特点及其带来的挑战,并且针对这些挑战,提出相应的策略。首先,采用Soft Balance Sampler来处理Intra-Domain imbalance和Inter-Domain imbalance问题。针对训练数据不足、泛化性难的问题,采用gridmask和autoaug来缓解。对于车载场景的小目标检测的挑战,采用Mosaic马赛克增强和Soft-nms来提高对小目标的检测效果。最后,基于anchor-base和anchor-free两类模型进行融合,有效提升了检测效果。

With the digital and intelligent development of the automotive industry, smart cockpits have become the future trend. This paper conducts research on the task of detecting leftover objects in the vehicle. First, we analyze the characteristics of the image data of the vehicle scene and the challenges it brings, then proposes corresponding strategies for these challenges. First, we use Soft Balance Sampler to deal with Intra-Domain imbalance and Inter-Domain imbalance. Aiming at the problems of insufficient training data and difficult generalization,gridmask and autoaug are used to alleviate the problems. For the challenge of small object detection in vehicle-mounted scenes, mosaic enhancement and Soft-nms are used to improve the detection effect of small targets. Finally,based on the fusion of anchor-based and anchor-free models, the detection effect is effectively improved.