Object Detection by Attention-Guided Feature Fusion Network

Abstract

One of the most noticeable characteristics of security issues is the prevalence of “Security Asymmetry”. The safety of production and even the lives of workers can be jeopardized if risk factors aren’t detected in time. Today, object detection technology plays a vital role in actual operating conditions. For the sake of warning danger and ensuring the work security, we propose the Attention-guided Feature Fusion Network method and apply it to the Helmet Detection in this paper. AFFN method, which is capable of reliably detecting objects of a wider range of sizes, outperforms previous methods with an mAP value of 85.3% and achieves an excellent result in helmet detection with an mAP value of 62.4%. From objects of finite sizes to a wider range of sizes, the proposed method achieves “symmetry” in the sense of detection.

Sparsity-Robust Feature Fusion for Vulnerable Road-User Detection with 4D Radar

“…The input of the SRFF, as shown in Figure 2, is the aggregated outputs of each backbone stage, each with a different scale. The SRFF module is inspired by the work of Shi et al [2], which follows a similar approach in the context of 2D object detection. However, our approach is specially adapted to the processing of sparse radar data.…”

“…These systems differ in terms of the sensor modality used. In scientific publications, the use of mono or stereo cameras is the most common [2], followed by 3D LiDAR sensors [3]. Both modalities show convincing results on scientific datasets under good weather conditions but are unfortunately limited in practical application.…”