<p>This work introduces a deep Encoder–Decoder network that employs a Multi-scale attention mechanism for fast and efficient road space detection using 3D LiDAR point clouds. The unstructured LiDAR data is used to generate colored Bird’s Eye View(BEV) images that offer vivid and a detailed representation of the road environment. With these generated images, the free space detection task is reduced to a 2D-scale problem that can easily be worked on with a fast and lightweight convolutional neural network. With our approach, fast inference is possible due to the reduced computational and memory access costs of the model, and to increase the detection accuracy, a Multi-scale attention module that is heedful to both spatial features and channels that are more informative to the task at hand is utilized in the Decoder. Furthermore, to mitigate the effects of working with small datasets, we propose a hybrid loss function that promotes a balance between precision and recall while preventing the over-fitting of our model. We demonstrate the efficacy of our proposed approach through qualitative and quantitative evaluations on the KITTI road dataset, and it indicates that the proposed ACNNet achieves excellent results compared to that of the models used for the comparative task even though our work does not make use of any pre-trained weights during training.</p>