Counting crowds on buses without slowing down the computer

Counting passengers accurately on public transport requires handling occlusion, perspective distortion, and wildly varying density. This work replaces conventional attention modules (which add computational weight) with parameter-free alternatives—PFCA, SA, and SimAM—that extract useful patterns without extra learnable parameters. Testing on CSRNet with the ShanghaiTech dataset, a hybrid approach (PFCASA) performed as well as or better than parameterized baselines, with different mechanisms excelling at different crowd densities. Result: the same accuracy, smaller model, faster inference on edge devices.