ACRerNet:Implement adaptive convolution in anchor-based lane detection

Accurate lane detection is crucial for autonomous vehicle safety, yet the slender, variable morphology of lane markings hampers robust feature learning. Conventional CNNs, with isotropic square kernels, fail to capture directional and long-range structural patterns, limiting performance in complex driving scenarios. To overcome these limitations, we present an Adaptive Convolutional Rerouting Detection Network (ACRerNet). In this network, we propose a deformable strip convolution module (KAConv-V2) that adaptively adjusts the aspect ratio of the convolution kernel through a lightweight parameter prediction network. This adaptivity ensures that the geometric shape of the convolution kernel can precisely match the elongated form of lane markings. To further enhance directional modeling, we propose a direction-aware convolution module (KAConv-V1) for real-time estimation of the optimal rotation angle of the convolution kernel, thereby ensuring that the receptive field is accurately aligned with the local lane direction and improving the representation capability of directional features. Finally, we present a channel attention (ISNet) that adaptively refines multi-scale features, effectively reducing background and occlusion interference. Comprehensive experiments on the public CULane and CurveLanes benchmarks confirm that our method achieves greater robustness in complex scenarios, surpassing existing mainstream approaches.