Human Action Recognition in Dark Videos

Image processing and action recognition in images are one of the most researched topics in Deep learning. Combining these two concepts for action recognition in lowlight footage is useful in a variety of applications, including night surveillance and self-driving at night. Due to the low photon count and SNR, video in low light is difficult. Short exposures videos are prone to noise, while long exposures can result in blur and are often impractical. To get a better understanding of the presented Action Recognition in Dark(ARID) dataset, which has low light videos divided into it’s action, making it an image classification problem. We examined it in depth and demonstrated it’s utility using simulated dark images. On this dataset, we also benchmarked the performance of existing action recognition models and investigated possible strategies for improving their performance. We introduce a novel pipeline for low-light images using RenNets and statistical image processing methods to identify the human’s actions in it to support the development of learning-based pipelines for human actions recognition in dark videos. We present promising findings from the latest dataset improving the top-1 accuracy by 3.8%. We also examined performance-related causes, and identify areas for potential research.