Action Recognition from Videos using Multi-Feature Trained Improved LinkNet Model

Video action recognition is a key application of computer vision (CV). Using data from prior observations, its primary objective is to accurately describe human behavior and interactions. The ability to recognize, understand, and predict complex human behaviors advances a number of important applications. In recent years, the CV community has given deep learning (DL) particular attention. Thus, utilizing DL technology, this work suggests a novel action recognition model in videos. First, the median filter model is used to pre-process the input frames following the video-to-frame conversion. Features such as the Median Binary Pattern (MBP) and Motion Boundary Scale-Invariant Feature Transform (MoBSIFT) are then extracted. Improved Link Net, which has better activation and loss functions, is presented for action detection after feature extraction. For TD=90%, the enhanced link net achieves a higher value of 0.9590. However, the accuracy results for LinkNet, DenseNet, LeNet, ResNet, SqueezeNet, and LSTM were lower, at 0.87329, 0.8795, 0.8708, 0.8944, 0.8795, and 0.884, respectively.