Machine Learning Classification Techniques for Human Activity Analysis Assistant Professor Mrs Anitha J1 Ramya R2 1 2 Professor of Dr Ambedkar Institute of Technology, Dept of MCA, Bangalore-560056, Karnataka, India Student of Dr Ambedkar Institute of Technology, Dept of MCA, Bangalore-560060, Karnataka, India Abstract —Human Activity Recognition (HAR) is for sure an enormous field of study and examination. This is a time series classification task in which data from multiple time steps is required to appropriately characterize the current activity. To put it differently, movement recognition is the most typical method of identifying or anticipating which movement or action will be taken up by someone. This is a series data classification problem in which data from multiple time steps is required to appropriately categorise the current activity. Eating, Walking, resting, reading the newspaper, conversing, jumping, standing, drinking, and sitting are examples of activities. The problem here is that to effectively forecast the action that is being performed, you will essentially need a succession of knowledgebased points in action recognition. As a result, the action-based recognition would be in the form of a statistic classification, with the disadvantage that the data from a sequence of time steps is likely to be insufficient to appropriately identify the action that is being performed. We used a variety of machine classification approaches to analyze human being behaviour, including, Neural Network, SGD, K-Nearest Neighbor (KNN), Naive Bayes & Random Forest. The results of the experiments showed that the Neural Network and logistic regression offer superior accuracy for human activity recognition compared to other classifiers including k-nearest neighbor (KNN), SGD, Random Forest, and Naive Bayes, although they require more processing time and memory. Keywords: Activity Recognition, Neural Network, Random Forest, SDN, k-nearest neighbor, Activity, Recognition, Classification. I. INTRODUCTION Human activity recognition (HAR) is an important exam question with accurate recognition of different pieces of training that can be stored in different ways. Human Activity Recognition has seen use in a variety of businesses and occupations, including smart healthcare, automation surveillance, and security. Computers have become capable of handling some very difficult jobs as a result of numerous breakthroughs in machine learning (such as comprehending an image). At the point when given a picture, models are being constructed that can foresee what's going on with the picture or distinguish if a specific trademark is there. These models depend on the system and working of the cerebrum and are known as brain network models (or fake brain organizations). Deep learning is a part of the directed discovery that looks at brain organizations, and various sorts of brain networks have been utilized to take care of an assortment of issues after some time. Human actions have a built-in hierarchical structure that displays the activity's varying degrees, which can be divided into three groups. At the most basic level, there seems to be an atomic element, and all these movement elements are also what help compensate for more complex human acts. Just after the action primitive level, comes the action/activity level. Ultimately, complicated connections, which propose human physical activities related to greater humans and things, are the nice level II. PROBLEM DEFINITION This project will most likely develop a model that can recognise basic human behaviours such as jogging, sitting, clapping, hand-waving, and boxing. A progression of recordings will be provided to the model, with an individual doing an activity in every film. The video's name will constitute the occasions which can be unfolding in this particular scene. Before it can examine the label of a supply of the information (videos) it has in no way seen, the computation needs to recognize this relationship. When shown a few models, the model should be able to detect various human demonstrations. Indeed, by claiming to show certain patterns, the model should be able to detect various human manifestations. III. EXISTING SYSTEM In controlled conditions, the bulk of known techniques shows human activities as a collection of picture elements captured in surveillance images or photos and various classification patterns are used to identify the main activity tag. The datasets considered are all generic. These limits, on the other hand, create an unrealistic environment that does not account for real-world conditions and fails to meet the requirements for an acceptable human activity dataset. Examine how people go about their daily lives, such as studying and eating. We can employ wearable device sensors to collect data from testing and training. After gathering the data, it will transmit it to the record output. IV. PROPOSED SYSTEM In this proposed system, we present the convolution neural network method for video action recognition. The input footage will be recorded using the webcam. The video sequence's number of frames is changed. The precise frame segment is subsequently identified by using CNN (Convolution Neural Network) approach. The maximum criteria weights from the extracting features frames are subsequently obtained using a convolutional neural network. Ultimately, the activity in the video will be identified, and a labelling (activity name) will be assigned. The output is then transferred to Firebase, where the user receives the Firebase value through an Android notification. A. Detection - The location of a person in a static photograph or a series of photographs, such as development photographs, is used to discern activity. The adaptability with an image can be followed in the case of a continuous grouping, although this is more important for applications like gesture communication. The fundamental reason for activity acknowledgement is that human vision can distinguish objects that robots can't with similar accuracy. It resembles a man mishandling his senses to find an item according to the viewpoint of a machine. B. Movement Recognition - The reason for Movement Recognition in registering science and specialized detail is to examine human activities utilizing numerical calculations and picture and camera tests. Gesture recognition techniques, on the other hand, are used to identify and recognise stance, movement, expressions, and human behaviours. The Neural Network is implemented using the Orange Tool, and thus the Neural Network is run in Python utilizing dropouts, yielding superior results. The data was processed using regression algorithms, K-nearest neighbours, SDG, Multilayer Perceptron, Naive Bayes, and Random Forest. The Recall and Precision Values of each model were determined, and the Confusion Matrix for each model was constructed. Fig 2: System Architecture Random Forest: Random Forest is a backslide and depiction segment. A wide range of trees fills in erratic timberlands. Each tree is produced utilizing the readiness information and a bootstrap test. Choose the amount of accommodating trees to be associated with backwoods (Numbers of trees within timberland region) and the number of properties to be drawn for thought at each middle point. If the not entirely set in stone (decision Number of left excessive), the number is comparable to a square base of the number of qualities in the data. The client can pick the profundity by which the trees will be developed since Unique Breiman's proposition will be to foster all trees without pre-pruning. Stochastic Gradient Descent: Stochastic Gradient Descent gadgets utilize stochastic slope plunge to limit a direct misfortune work. Choose the calculation parameters for the coordinating goof work, backslide incidence limit, and guidelines to avoid overfitting. Finally, the learning limits are set, and they are not permanently set up by the learning rate. For consistent cycles, the beginning readiness botch is 0.0100, and the number of overlooks the scheduling data is fixed at 1000. Programming Language: Python is feasible in all circumstances, particularly for AI and AI estimations. The problem is that the models don't have complete trust in each frame buffer prediction in every case, thus these assumptions will change extremely quickly and without error because the calculation doesn't witness the actual video transfer. One of the fundamental solutions to this sort of trouble is to restrict the casings instead of parting and showing the results of a solitary structure. To obtain this result, after the valuation of n, use anything very basic like a moderate movement or moving in the centre. Fig 1: Human Activity Recognition C. Design Approach – Since the data offers a multivariate classification issue, we used both supervised and unsupervised algorithms. Naive Bayes: Naive Bayes is one probabilistic model that considers speedier expectations from information continuously. KNN: The KNN is a solicitation approach that requires no past information on the information course. Assign a value of k to the number of nearest neighbours. The unidentified datapoint will be given out to a class from a get-together of named places in its close region. The distance equation, which is either Manhattan or Euclidean, is used here. The KNN device utilizes a calculation that searches out the k nearest planning model in featured space and uses their normal as a gauge. At the point when k is 5, the KNN calculation gives 96.6 per cent exactness in following human exercises. D. Working - Videos are organised into action categories on the internet, including cricket, piercing, cycling, and so on. This database is frequently utilised in the creation of action viewers, a video sharing software. Frames in a video are in a specific order. Keras is a Python package for open-source development and testing of in-depth learning models that is very effective and simple to use. It combines the Theano & TensorFlow frameworks to let you design and build models with just a few program codes. E. System Implementation Hardware Requirements : • • • Intel Core i3 3rd gen processor or later. 512 MB disk space. 512 MB RAM. Software Requirements : • • • • Microsoft Windows XP or later / Ubuntu 12.0 LTS. or later /MAC OS 10.1 or later. Python Interpreter (3.6). TensorFlow framework, Keres API. PyQT5, Tkinter module. V. RESULTS The camera first detects the presence of a human, then the action that is occurring. The deep learning algorithm utilised is ResNet, which generally employs 2D kernels, however, 3D kernels are enabled to improve the effectiveness of activity recognition. The kinetics400 dataset, which specifies 400 various types of work, was used to train the activity recognition model. In addition, 400 videos for every action are available to improve accuracy. For activity identification, the ResNet network for human action recognition with 2D CNN employs 152 layers. This 2D CNN ResNet network handles tasks such as detection, segmentation, and captioning, but our model additionally includes a 3D CNN that processes videos rather than images, increasing the likelihood of action recognition success. Recognition, summarization, optical flow, segmentation, and captioning are some of the activities that can be completed. As a result, human activity recognition accuracy is improved. VI. CONCLUSION We have suggested a machine learning-based Human Action Recognition system that deals with identifying activity as regular or suspicious based on its nature. If any unusual activity is detected, an immediate alert signal is issued to authority, allowing for the reduction of further depressing repercussions. Recognizing human hobbies is a vast area of study that entails looking at such objectives to comprehend someone's genuine motion or movement. This is a type of time collection sort problem in which statistics from a series of time steps are required to properly classify the current movement. Recognizing human interest, or HAR is a difficult task of classifying statistical elements. So, for future work, we will run the model using more live datasets and get the result. We'll also look into different perceptron and optimizers, as well as more advanced ways to improve model accuracy. Future work may also include the implementation of this model for useful applications like monitoring the activities, predicting danger, etc. VII. FUTURE SCOPE Human Activity Recognition can help with home automation surveillance, health services, security surveillance, and childcare, among other things. We may upgrade our programme in the future by employing object activity recognition, which allows us to track and analyse the activities that objects conduct. The use of combined huge datasets can be used to identify activities occurring at a lower rate of time. The system should be able to detect even the tiniest of differences. If the actor doing the anomalous action is not captured in the first place, the data of the actor can be stored and the actor identified. Activities that repeat themselves should be saved to save time and space throughout the recognition process. Such a paradigm can likewise be implemented in the government authority sector. Much more work can be done in terms of improving precision and dealing with difficulties such as optical identification and image background clutter. Human activity recognition is mostly employed in health systems deployed in residential settings, hospitals, and rehabilitation facilities. It's widely utilised in rehabilitation programmes for chronic disease management and prevention to monitor the activities of older individuals. REFERENCES [1] Amit Konar, & Aruna Chakraborty. (2015). Emotion recognition : a pattern analysis approach. John Wiley & Sons, Inc. [2] Md Saad, M. H. et al. (2011) “Event description from video stream for anomalous human activity and behaviour detection,” in 2011 IEEE 7th International Colloquium on Signal Processing and its Applications. IEEE, pp. 503–506. [3] Simonyan, K. and Zisserman, A. (2014) “Two-stream Convolutional Networks for action recognition in videos,” arXiv [cs.CV]. Available at: http://arxiv.org/abs/1406.2199. [4] Tran, D. et al. (2017) “ConvNet architecture search for spatiotemporal feature learning,” arXiv [cs.CV]. Available at: http://arxiv.org/abs/1708.05038. [5] Alsheikh, M. A. et al. (2015) “Deep activity recognition models with triaxial accelerometers,” arXiv [cs.LG]. Available at: http://arxiv.org/abs/1511.04664. [6] Anguita, D. et al. (2012) “Human Activity Recognition on Smartphones Using a Multiclass Hardware- Friendly Support Vector Machine,” Springer International Workshop on Ambient Assisted Living.Lecture notes in Computer Science, 7657, pp. 216–223. [7] Ali, S. and Shah, M. (2007) “A lagrangian particle dynamics approach for crowd flow segmentation and stability analysis. In Computer Vision and Pattern Recognition,” in CVPR’07. IEEE Conference on. IEEE, pp. 1–6. [8] Agarwal, I., Kushwaha, A. K. S. and Srivastava, R. (2015) “Weighted Fast Dynamic Time Warping based multi-view human activity recognition using a RGB-D sensor,” in 2015 Fifth National Conference on Computer Vision, Pattern Recognition, Image Processing and Graphics (NCVPRIPG). IEEE, pp. 1–4. [9] Hossain Shuvo, M. M. et al. (2020) “A hybrid approach for human activity recognition with support vector machine and 1D convolutional neural network,” in 2020 IEEE Applied Imagery Pattern Recognition Workshop (AIPR). IEEE. [10] “SOM-Based Data Visualization Methods. Intelligent Data Analysis, Laboratory of Computer and Information Science” (1999) Finlandvol, 5400(2), pp. 111–126.