An Efficient Dense Pedestrian Detection Model Based on RT-DETR

Pedestrian detection is a vital study field within object detection. With the development of deep learning, models perform exceptionally well in pedestrian detection for sparse scenarios. However, their performance often fails in complex and crowded dense pedestrian scenarios, particularly when there is a concentration of small objects. Challenges such as low detection accuracy, high miss rates, and high false-positive rates persist in these scenarios. To enhance dense pedestrian detection performance, we propose a model named RT-DETR-LKG, based on RT-DETR. The improved model incorporates LSKA (Large Separable Kernel Attention) and the GELAN (Generalized Efficient Layer Aggregation Network) modules (LKG). The backbone network integrates the LSKA, which generates spatial attention weights for feature maps via multiple convolutional layers. This process enables feature extraction across various receptive fields and applies weighting to input channels, resulting in more refined feature extraction and enhancement. The hybrid encoder network applies the GENLAN module. It achieves rich feature extraction with minimal computational cost by combining channel partitioning and repeated convolutions. The network's feature extraction ability is also kept. In bounding box regression, the WIoU loss function is used instead of the GIoU loss function to improve anchor selection for occluded and small object prediction boxes. Experimental results indicate that the new RT-DETR algorithm performs better in crowded pedestrian scenarios. It has a mean average precision of 86.1%, which is 1.2% higher than the original model.