Human Detection in Crowded Scenes Using Hybrid ResNet
Bandar M. Alghamdi · 2023
Human identification in crowded places is challenging because of the overlapping and obscuring of things. Research on crowd management and surveillance techniques is critical for the protection of the general public. The development of an efficient CMS is complicated by issues including density fluctuations, unequal distribution of objects, occlusions, posture estimation, etc. The usefulness of human posture much exceeds that of widespread vehicle detection. To address the challenge of human object recognition in congested and overlapping environments, we present a hybrid ResNet, a realtime human identification network. By adding a deep cascade fusion module, the anchor-free model is able to better extract features from sparse settings including small objects. The centre region's prediction score is improved by the activation module's use of low occlusion features. In crowded areas, the network has an easier time differentiating between persons despite their changing postures. When it comes to realtime inference, our hybrid ResNet is virtually as quick as previous approaches, and it achieves competitive performance on the Crowd Human dataset.