Article Preview
Top1. Introduction
The health problem is one of the major issues that are becoming dominant as a result of lifestyle and various other disorders. Thus, to prevent such disorders, regular exercises can prove to be beneficial but inadequate knowledge about different types of exercise, the quantity of calories burned during physical exercise, and the correct method of performing them is a challenging task. To overcome this issue, an efficient trainer is required who can assist and correct the users at the wrong places. However, due to a lack of time and an addiction to the digital world, individuals currently rely more on smartphone applications, which can be accessed from anywhere with ease as per their requirements. This has originated a novel idea and motivated us to design a personal virtual trainer application that can recognize exercise in videos and enable information access flexibility. However, exercise recognition in videos is a challenging task due to a variety of reasons, some of which are described below:
- ●
Exercises range from simple movements of the arms or legs to more sophisticated movements involving several body parts.
- ●
Combination of more than one move named as different exercise. To handle it, a memory-based learning model is required that identifies exercise based on multiple moves in a video.
- ●
The change in background, pose of the user, different illumination conditions, and various other factors make this task more challenging.
The applicability of action recognitions in multiple domains like entertainment (Kong et al., 2018), robotics (Chrungoo et al., 2014), computer vision (Zhang et al., 2017), sports (Soomro et al., 2014), and surveillance (Jin et al., 2015 ; Jin et al., 2018) has made it still an active area of research. Although several solutions for action recognition exist in the literature such as space-time features (Blank et al., 2005; Batra et al., 2005), trajectories (Wang et al., 2011), Support Vector Machine (SVM) (Schuldt et al., 2004), KNN (Paul et al., 2015), etc. but each method has its own set of limitations. Owing to the difficulties in processing video data depicting human activity, a novel study is proposed that uses CNN (Karpathy et al., 2014; Taneja et al., 2019; Choudhary et al., 2021) and bidirectional LSTM (Hei et al., 2015) to analyze and process the video data. The proposed method has the capability to process long and complex videos making it more efficient than baseline methods.
The key contributions of this research paper are as follows:
- ●
To develop a personal virtual trainer application that can assist the users while performing the exercise itself by recognition of their respective actions.
- ●
To provide the number of calories burnt by the user based on the physical exercise.
- ●
To provide detailed information of exercise along with the guidance of correct posture like the type of exercise, the progress in difficulty level of exercise.
- ●
To enrich the user’s entire experience at their ease of access.
- ●
To design a user –interactive smartphone application with control in the hands of the users.
Thus, the main objective of this paper is to help the users in staying fit by providing detailed information about the exercise along with the correct method of performing exercises. It facilitates users with the additional information that a user is desirous to know after performing exercises like the number of calories burned, type of exercise to be performed as per body requirements, and the progress in difficulty level of exercises. The above-mentioned information is gathered by learning the features of the user performing the exercise. In contrast to existing applications that only display a predefined set of exercises one after another without allowing users control, this study adds to the existing literature by giving users control as an option so that users can opt for exercises of their own choice (Gay et al., 2012). The novelty of the paper is that it puts the control in the hands of the users, enhancing the entire user experience.