A Coord-Aware Transformer for Pedestrian Fall Detection

The detection of fallen pedestrians is one of the key problems in automatic driving. Due to complex road conditions, the detection of fallen pedestrians during driving is still challenging for occlusion and small targets. Usually, existing methods such as convolutional networks have low detection accuracy for these cases. Therefore, in this paper we propose to apply the Coord-Aware Transformer YOLOv7 (T-YOLOv7) method to the detection of fallen pedestrians. The CoordConv Feature Pyramid Networks (C-FPN) and the CoordConv perception Detect (C-Detect) are designed. By introducing feature coordinate channels, the network can perceive the spatial information of the feature map, improve the detection of occluded targets. Add the Biformer module to the C-FPN to improve the detection accuracy of small targets. In addition, we use the Wise-IoU to replace the original network CIoU, it can focus on ordinary prediction boxes. The improved network model achieved 87.45
^%
mAP on the open source data set Fall detection Database, which increased by 3.72%mAP and 5.25
^%
mAP compared with YOLOv7 and PDD-FCOS, respectively.