An Efficient Low‐Complexity Deep Learning Model for Low‐Light Human Action Recognition Using IoT Devices

Human action recognition under dark and low‐light conditions is challenging because important spatial features are often weak and difficult to distinguish. This paper presents a lightweight Deep Learning (DL) model that combines a frozen MobileNetV2 backbone with a convolutional refinement head to improve the representation of spatial features before global pooling. The refinement head consists of two convolutional layers (Conv2D‐128 and Conv2D‐64), each followed by Batch Normalization and Dropout. L2 regularization and staged dropout reduce overfitting, decreasing the train‐validation gap from about 15% to below 5%. Evaluated on the ARID v1.5 dataset, the proposed low‐complexity deep learning model achieves an accuracy of 86.59% using only 3.81 M trainable parameters. Two simple parameter‐free inference techniques, frame clipping and late fusion, further improve the accuracy to 93.49%, surpassing a CNN + LSTM baseline (92.77%) without increasing model complexity. These results demonstrate that preserving spatial information and improving frame quality are more effective than adding temporal modeling for this dataset, making the proposed deep learning model a practical solution for IoT and edge‐based low‐light human action recognition.