Object Detection and Heading Forecasting by fusing Raw Radar Data using Cross Attention

Radar has been believed to be an inevitable sensor for advanced driver assistance systems (ADAS) for decades. Along with providing robust range, angle and velocity measurements, it is also cost-effective. Hence, radar is expected to play a big role in the next generation ADAS. In this paper, we propose a neural network for object detection and heading forecasting based on radar by fusing three raw radar channels with a cross-attention mechanism. We also introduce an improved ground truth augmentation method based on Bivariate norm, which represents the object labels in a more realistic form for radar measurements. Our results show 5% better mAP compared to state-of-the-art methods. To the best of our knowledge, this is the first attempt in the radar field, where cross-attention is utilized for object detection and heading forecasting without the use of object tracking and association.