EventASEG: An Event-Based Asynchronous Segmentation of Road With Likelihood Attention

Road detection is essential in autonomous driving, but conventional camera-based road segmentation is resource-intensive and slow. The emergence of event cameras presents a promising solution, with benefits such as low power consumption, data sparsity, and high temporal resolution etc. Existing approaches implement synchronous frame-based encoder-decoder architectures, resulting in significant computational overhead. Hence, we introduce EventASEG (Event-based Asynchronous Segmentation Network), a lightweight decoder-less transformer-based architecture, which significantly reduces the computational load. By directly processing the asynchronous and sparse event data, EventASEG eliminates the overhead of dense processing and event-to-frame conversion. However, processing raw events is challenging due to the limited contextual information of the scene. To effectively handle this lacuna, we introduce an innovative likelihood-based self-attention module. Comparative studies demonstrate that EventASEG achieves state-of-the-art performance on public datasets and operates with low latency and computing. Our approach is optimized for binary segmentation tasks and is particularly well-suited for time-critical and power-constrained robotic platforms.