An Omnidirectional Visual Perception Method Based on Parameter-Free Projection of Key Points
Mingchi Feng, Yi Chen, Haoyu Bai, Kun Wang · 2024
In recent years, omnidirectional vision BEV (Bird's Eye View) perception has become an important research direction in the field of autonomous driving vision. The correct perception of instance targets around the vehicle is a key technology to ensure that the vehicle can drive safely in complex traffic scenes. The difficulty of this technology is to efficiently convert the 2D features of the camera view into BEV features. Existing methods based on depth estimation suffer from the problem of inaccurate depth estimation or the need to create a large number of BEV depth labels. The model size of the Transformer-based method is large, which makes it difficult to use in practical applications. Both methods have the problem of a large number of invalid calculations, which seriously affects the real-time performance of the algorithm. To solve this problem, in this paper, we propose a non-parameterized key point projection method based on the attention mechanism, which significantly improves the computational speed of the algorithm. Our IOU for instance segmentation on the NuScenes is 44.4, and the algorithm runs at 65 FPS. Experiments show that the segmentation performance strength of our method outperforms that of similar methods and meets the real-time requirements.