Radar object detection with cross‑scale feature fusion and adaptive temporal manifold convolution

Millimeter-wave radar is widely regarded as the key technology for object detection in complex environments. However, its actual performance is often limited by two inherent factors, low radar spatial resolution and strong timing dependence between continuous radar frames. In response to these problems, we have developed a new radar object detection method, which combines cross-scale feature fusion and adaptive temporal manifold convolution. In order to solve the challenges of spatial resolution, our model introduces a cross-resolution feature fusion module called CrossUPP in the encoder-decoder architecture. The module establishes an upsampling skip connection between the deeper encoder layer and the shallower decoder layer, which significantly improves the ability of the network to detect objects at different scales. In order to capture time dynamics, we have introduced an adaptive temporal manifold convolution module. The module aims to take advantage of the continuity between frames and the subtle local geometric changes in the radar sequence. By capturing these time clues, the module enhances the feature representation of the network and improves its time modeling ability. Furthermore, we applied a simple constraint mechanism in the post-processing process to further optimize the detection results. We evaluated our approach on the ROD2021 dataset. Our CrossUPP-based model is better than the representative baseline model RODNet. The average accuracy average (mAP) has increased by about 5.74%, and the average recall rate average (mAR) has increased by about 7.07%.