Dual-stream infrared object detection network combined with saliency images

Qin, Peng; Liu, Yunfeng; Zhou, Gaofan; Tang, Chuanming; Zhang, Jianlin

doi:10.1109/ipec54454.2022.9777325

Cited by 2 publications

(1 citation statement)

References 16 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Zhewei Xu aggregated ground plane context to Faster-RCNN and constructed a GPCAnet network (GAP + GRPN + Backbone), reducing FP cases [14].Minglei Li modified the YOLOv3 algorithm with MobileNetV2 for the detection of infrared and visible images to improve the MAP of pedestrian detection [15]. Peng Qin used CNN and Transformer fusion networks for infrared and pseudo-color image detection to enhance the precision [16]. Jonathan Kim used two EfficientNet backbone networks to extract features on the KAIST dataset.…”

Section: Introductionmentioning

confidence: 99%

Pedestrian detection based on improved EfficientDet

Zhang,

Yu,

Min

et al. 2023

Preprint

View full text Add to dashboard Cite

Pedestrian detection plays a critical role in computer vision applications, but the performance of existing algorithms utilizing visible images is inadequate when confronted with poor lighting conditions. Conversely, infrared images deliver superior outcomes in low-light environments. Based on this issue, this study proposes the Double-EfficientDet-Improved (DEDI), a dual-stream model that integrates both visible and infrared imagery by enhancing the EfficientDet network structure. Firstly, the shuffle module is integrated into the backbone network to facilitate inter-channel information exchange within the feature layers. Secondly, by controlling the gradient path of the backbone network, the feature extraction ability is improved and the robustness is stronger. Lastly, a Multi-scale Fusion Module (MFM) is incorporated into the proposed model to harmonize information from both visible and infrared images. Experimental results demonstrate that this DEDI method achieves higher mean average precision (MAP) scores of 74.72\% on the KAIST dataset and 95.13\% on the LLVIP dataset, which is better than other object detection methods.

show abstract

Section: Introductionmentioning

confidence: 99%