Recognising Actions with Segmentation and Prediction Techniques in ROI based Deep Learning Framework

Authors

  • Manoj Kumar .K, L. Sujihelen

DOI:

https://doi.org/10.17762/msea.v71i4.977

Abstract

Predicting the action of human beings has been an interesting research domain in the recent decades. Computer Vision is the domain associated with monitoring of video sequences for extracting meaningful patterns and predicting the future action by mapping the features of one or more frames. Motion detection, featured patterns derivations and supposed action prediction are defined by annotations and human action labelling performed through conventional neural network architectures. The known shortcoming of the conventional techniques is the exclusion of temporal features from annotations. Considering the cost associated with human action prediction, many models tend to exclude the temporal features or in certain models, the feature cannot be included into the mutual processing model along with other significant features. The proposed model investigates the challenges and difficulties associated with traditional models, and hence provides a resolution to explore the betterments introduced in the pre-processing stages. Regions of Interests are defined in potential frames of video sequences to highlight the important features and these ROI will be acting as representative factors for deriving better outcomes. The proposed model demands lesser computational costs as the number of features to be used in computation depends on ROIs, thereby limiting the number of required resources. The ROIs are described using the timestamps (start and end) apart from the location of those regions, enabling them to determine the actions based on action queries. The proposed model is categorized into pre-processing with background separation and HOG, followed by censoring of relevant regions of interests, summarizing the action frames, and finally predicting the future actions. Since the temporal information is considered in the proposed approach, action based queries and prediction of future action is facilitated by proper action based video segmentation. Investigated simulation results prove that the proposed technique predicts the complete actions with greater accuracy and quicker detection, better than conventional techniques.

Downloads

Published

2022-09-28

How to Cite

Manoj Kumar .K, L. Sujihelen. (2022). Recognising Actions with Segmentation and Prediction Techniques in ROI based Deep Learning Framework. Mathematical Statistician and Engineering Applications, 71(4), 4072–4090. https://doi.org/10.17762/msea.v71i4.977

Issue

Section

Articles