Abstract:

Detecting rare objects in transportation environments is crucial for the safety and reliability of autonomous driving systems. To address the challenges of limited detection capability on rare objects that appear only a few times during training, we propose an improved long-tailed object detection method for traffic scene images named TF-DETR (tail free DETR) based on RT-DETR model. First, the lightweight CSP-EDA (CSP-efficient depthwise aggregation) was constructed toreconstruct the backbone network. Its core unit, the efficient depthwise aggregation, integrated the pooled-transpose attention and convolutional gated linear unit to enhance feature representations for tail-class samples. Second,the spatially-adaptive dynamic cognitive interaction was deployed in the encoder. Through feature reuse between the two-stage, the interaction module established an efficient information interaction mechanism. The mechanism spanned from feature extraction to multi-scale cognitive enhancement for tail-class objects. Finally, the pinwheel-shaped convolution wasintroduced for feature map downsampling. Through its collaborative design of asymmetric padding and multi-directional kernels, the model enhanced detection capability for tail-class objects while maintaining a lightweight architecture. Experimental results on the CODA2022 dataset show that TF-DETR achieves a 4.9% increase in AP50 and a 5.4% increase in AR50 compared with the baseline. Meanwhile, a 22.9% reduction in FLOPs and a 25.7% decrease in Parameters are realized.

Key words: transportation scene images, object detection, deep learning, long-tail distribution, RT-DETR