Crowd counting and localization, fundamental computer vision tasks, hold significant promise for applications in public safety and intelligent retail. However, the performance of existing methods degrades drastically when faced with challenging real-world scenarios, particularly under complex indoor lighting conditions. This is primarily attributed to their reliance on fixed, hand-crafted representations, which cannot adapt to the dramatic appearance changes caused by glare and shadows. To address this issue, we propose a novel model named AFT-Net. At its core is an Adaptive Focal Inverse Distance Transform (AFT) mechanism, which dynamically learns the optimal representation function for individual locations via a Hyper-Network. Furthermore, we incorporate the Convolutional Block Attention Module (CBAM) to enhance the network’s perception of critical regions. A key feature of our design is that the AFT module is only active during training to guide the learning process; it is discarded during inference, ensuring no extra computational cost. To systematically evaluate this problem, we construct and release a new Indoor-Robustness Benchmark dataset. Extensive experiments demonstrate that AFT-Net achieves state-of-the-art (SOTA) performance and superior robustness, fully validating its effectiveness.