Abstract:
One of the fundamental tasks in autonomous driving is lane line detection. We aim to improve detection accuracy in complex scenarios and strike a better balance between performance and complexity of lane line detection networks. To this end, we first propose Parallel Spatial Separation Convolution (PSS-conv), a new convolution operation built on a new parallel spatial convolution decomposition and a channel-weighted feature merging strategy, to aggregate the features obtained from decomposed convolution. Then, we propose Parallel Spatial Separation Convolution with Message-Passing (PSSconv-MP), in which a new message passing module is added before feature merging to enable slice-by-slice information propagation. Based on the PSS-conv, PSS-conv-MP and residual connection, we construct a new lane line detection network called Parallel Spatial Separation Network (PSSNet), which can handle challenging scenes like curve and obscured lane lines. Extensive experiments show that PSSNet can achieve a superior performance on the challenging lane line detection benchmark CULane.
