In order to improve the accuracy of road detection, a roadside lidar drivable space detection method based on an improved fully convolutional neural network (FCN) was proposed. Firstly, a two-dimensional top view is generated by building a ring grid map and counting the point cloud information in the grid. Then, a hybrid dilated convolution is introduced in FCN to replace the standard convolution, and a spatial feature alignment module is added after the pooling layer, and a channel feature alignment module is added after the convolution layer to construct HCS-FCN. HCS-FCN is compared with traditional FCN and SegNet on the self-constructed 16-line roadside lidar road dataset and 32-line roadside lidar road dataset. The experimental results show that the F₁-score of HCS-FCN reaches 88.4% and 89.2% on the 16-line dataset and 32-line dataset respectively; the average pixel accuracy reaches 89.1% and 89.7% on the 16-line dataset and 32-line dataset respectively; the average pixel intersection over union ratio reaches 87.3% and 88.9% on the 16-line dataset and 32-line dataset respectively; all of which are better than traditional FCN and SegNet.