HiBi-RTDETR: an efficient real-time infrared traffic object detector for edge devices

To address the challenges inherent in infrared road traffic object detection, such as the low contrast, high noise, and blurred details characteristic of infrared imagery, alongside the stringent real-time processing demands of traffic scenarios, we propose the HiBi-RTDETR detector, an enhanced model built upon the RT-DETR (Real-Time Detection Transformer) framework. The model incorporates two core innovations: First, we integrate the HiLo attention mechanism (High-Low Frequency Attention), which enhances the model's perceptual accuracy for infrared images by disentangling and processing high-frequency local details and low-frequency global structural information separately. Second, we employ a Bidirectional Feature Pyramid Network (BiFPN) to optimize feature fusion, while innovatively incorporating S2 features to bolster the model's multi-scale detection capabilities. To further satisfy the real-time requirements of road traffic applications, we propose a Sorted Wavelet Channel Pruning module. This technique compresses the model by selectively retaining channels that contribute significantly to edge information while pruning those with lesser contributions. Experimental results on the CTIR infrared dataset demonstrate that the pruned HiBi-RTDETR model achieves a 50% reduction in computational complexity (FLOPs) with only a marginal 1% decrease in mean Average Precision ([email protected]). Crucially, the optimized model attains an inference speed of 33.2 FPS on the NVIDIA Jetson Nano embedded platform, representing a 26.3% improvement over the pre-pruning baseline and successfully meeting the real-time detection threshold.