Human Recognition Using Deep Learning Techniques

Introduction

The advancement of technology overtime has led to automation of tasks and development of algorithms that are adaptive to change and as a result can engage in a wide range of complex activities. Vadapalli (2020) highlight that Machine Learning and Artificial Intelligence have significantly changed the world over the past few years with breakthrough innovations that have been adopted in various fields including manufacturing and processing industries, education, business and security. In security particularly, the development and application of sensors and cameras in surveillance has significantly improved the level of security afforded to both private and public establishments. With the development of Deep Learning Techniques consisting of three types of layers including input, hidden layer and output similar to the human brains neural network, machine learning has been elevated to a whole new level where machines can identify characteristics and recognize patterns and tasks (Altexsoft, 2019). Image recognition is among the tasks in which Deep neural networks (DNNs) adopted in deep learning techniques excel, and with the current emergence of Human Activity Recognition (HAR) protocols as a result of the cutting edge technology (Muralidharan et al., 2021), this study looks to evaluate whether deep learning techniques can be adopted to recognize abnormal human behavior in front of a surveillance camera.

Problem Statement

The advancements of technology has sure inspired new and much more effective and efficient ways to engage in tasks. For instance, automating repetitive processes, and taking up more complex human related processes due to the adaptive nature of AI and machine learning and summarily making machines adopted for various activities even better than they were before due to the adaptability component. However, while this adaptability enhances various human activities such as security, making organizations and individuals capable of better protecting themselves and their property, criminals and criminal activity have also been impacted by this same technology. Criminal enterprises are capable of hiding in plain sight and carrying out criminal activities without recognition either by hacking into systems or using the very systems designed to identify them to conceal themselves and prevent them from being caught. As such, it is necessary to develop much better and proactive ways of identifying malicious individuals and any malicious activities before hand so as to maintain the high levels of security required by security firms and personnel changed with the protection of both public and privatized assets. This research as such sets forward to investigate whether human recognition using deep learning techniques can be adopted to recognize abnormal human behavior in front of surveillance cameras.

Aims and Objective

The study aims to determine whether human recognition using deep learning techniques can be adopted to recognize abnormal human behavior patterns through surveillance cameras and whether the concept can be applied to actively identify malicious or criminal individuals and activities within both public and privatized establishments. If possible this can be an additional high technology security feature for security companies to enhance better protection of their assigned establishments.

Research Objective

In the achievement of the aim, the study will be divided into three major and precise objectives.

To critically evaluate the literature regarding deep learning techniques and how they relate to deep neural networks

To establish how deep learning techniques are used and whether they are efficient in the process of human activity recognition

To find out whether deep learning techniques can be adopted to recognize abnormal human behavior patterns through surveillance cameras

Research Questions

How does deep learning techniques relate to the human brains functioning and deep neural networks?

In what ways are deep learning techniques adopted to identify distinct human activity and their implications?

Can deep learning techniques be used to recognize abnormal human behavior patterns through surveillance cameras?

Rationale / Intended User

While studies and research behind human activity recognition have picked up in the recent past especially with the development of machine learning and artificial intelligence (Ravi et al, 2016; Vrigkas, Nikou and Kakadiaris, 2015: Brownlee, 2018; Jobanputra, Bavishi and Doshi, 2019; Erdas and Guney, 2021 ), a study in the determination of their efficacy for use within the crimes and security departments is yet to be carried out and as a result, security firms and companies are yet to adopt its use. The outcomes of this study will be able to not only confirm or de-confirm the efficacy and efficiency of using deep learning techniques for human activity recognition through surveillance cameras, but it will also advance critical and understandable knowledge on how deep learning techniques are used in the identification of abnormal human activity through surveillance cameras as well as highlight why this process is similar to the process adopted by a human brains’ neural network. The outcomes of this study if found to be adequate and effective will enable the adoption of deep learning techniques to identify abnormal and malicious human behavior via surveillance cameras for security companies as well as the governments criminal investigations departments. This will further enable enhanced security of public and privatized establishments.

System Requirements, Project Deliverables and Expected Outcomes

The final product should include the implementation of a deep learning technique in the identification of abnormal human behavior through surveillance cameras. This includes visual software that is able to analyze different visual cues in the behavior of humans from visual data such as camera feeds and determine when an individual is exhibiting abnormal behavior. The project deliverables will include the running of multiple simulations both including malicious and criminal individuals behaving in ways that can be considered abnormal and others that have no malicious individuals in sight. The research will as such look to find out whether through the adoption of multiple deep learning techniques, the software can identify such malicious and abnormal behavior from the individuals within the visual simulations presented. The expected outcome is that abnormal human activity recognition is possible with the use of deep learning techniques. The resultant software developed will be able to identify abnormal behavior patterns presented by individuals that could suggest malicious or criminal intent with a considerable degree of accuracy (above 60%).

Primary Research Plan

The study will involve a software algorithm with deep learning techniques capable of identifying visual cues from visual data such as surveillance camera footage. The software will be adopted in various simulations of criminal and non criminal activities to determine whether it can recognize any differences in the behavior exhibited by malicious individuals within the simulation data compared to other individuals with no malicious intent. The research will involve running multiple simulations of different data sets including visual recordings of previous criminal activities recorded such as bank robberies and shop lifters as well as live feeds of individuals acting out these scenes. Particularly, the database to be used is the ETH dataset which contains extensive visual data. The data collected will then be averaged to identify the percentage accuracy of the resultant software and whether it can be adequately adopted in a security and criminal setting. The study should generally take a day as the key variable being investigated is the machines response and whether the response is accurate in identifying abnormal human behavior.

According to Hu (2018) Eth is a dataset for pedestrian detection compiled from the stereo rig mounted on cars. The testing set of the dataset contains 1,804 images in three video clips with a resolution of 640 x 480. This dataset provides significant images by which to effectively test the accuracy and effectiveness of the software being developed. The software testing will also adopt auto encoder techniques for better results. Auto encoders are a type of neural network technique where input is ultimately the same as output. The system engages in actively compressing the input into lower dimension code and then reconstructing it to an output based on the inputs representation (Dertat, 2017). According to Jordan (2018) an auto encoder leverages the neural network for the task of representation learning. Compression and subsequent reconstruction of the output is often difficult for images and input features that are independent from one another. However if a sort of a structure or pattern exists within the data then it can be effectively learned and leveraged in the reconstruction of compressed input into similar output.

Literature Review

Deep Learning Techniques

Deep learning techniques are based on the deep neural networks technology which are basically computer systems designed to recognize patterns. These patterns may be audio or visual or in the form of data and as such the neural networks are developed based on the human brain structure with three distinct layers including the input, hidden layers and output. According to Altexsoft (2019) the input layer receives signals which are them processed within the hidden layer after which an output layer makes a decision or a forecast regarding the input data. While traditional neural networks have up to three hidden layers, deep networks that adopt deep learning techniques exhibit up to hundreds of hidden layers which allow them to process and make predictions about much more complex data such as images of the human face and human activities.

Similar to the brain which takes extensive and significant training to learn to recognize what is familiar and therefore safe and tell it apart from what is alien and potentially dangerous, deep learning models also require training to be able to the different characteristics to be used in identifying and making informed judgments and prediction about whether the behavior exhibited is natural or abnormal. Vadapalli (2020) highlights up to ten different deep learning techniques that can be used in the analysis of different types of data from audio to visual to alphanumeric data and can be applicable to multiple fields. Two of the major techniques that will be focused on within this study include

Convolutional Neural Networks – This is an advanced and high-potential type of artificial neural network model developed for handling highly complex data compilations. It is developed from the perceived order of arrangement of neurons present within the visual context of an animal’s brain and is considered among the most flexible model for specializing image data.

Generative Adversarial Networks – this involves a combination of two deep learning techniques of neural networks in a generator network which yields artificial data and a discriminator network tasked with helping to discern between the real and false data. Vadapalli (2020) asserts that it works best in image and text generation, enhancement and recognition.

References

AltexSoft, 2019. Image Recognition with Deep Neural Networks and its Use Cases. [online] AltexSoft. Available at: [Accessed 19 October 2021].

Brownlee, J., 2018. Deep Learning Models for Human Activity Recognition. [online] Machine Learning Mastery. Available at: [Accessed 19 October 2021].

Dertat, A., 2017. Applied Deep Learning - Part 3: Autoencoders. [online] Medium. Available at: [Accessed 27 October 2021].

Erdaş, Ç. and Güney, S., 2021. Human Activity Recognition by Using Different Deep Learning Approaches for Wearable Sensors. Neural Processing Letters, 53(3), pp.1795-1809.

Hu, Z., 2018. Pedestrian detection on ETH data set with Faster R-CNN. [online] Medium. Available at: [Accessed 27 October 2021].

Jobanputra, C., Bavishi, J. and Doshi, N., 2019. Human Activity Recognition: A Survey. Procedia Computer Science, 155, pp.698-703.

Jordan, J., 2018. Introduction to autoencoders.. [online] Jeremy Jordan. Available at: [Accessed 27 October 2021].

Muralidharan, K., Ramesh, A., G, R., Prem, S., A A, R. and Gopinath, D., 2021. 1D Convolution approach to human activity recognition using sensor data and comparison with machine learning algorithms. International Journal of Cognitive Computing in Engineering, 2, pp.130-143.

Ravi, D., Wong, C., Lo, B. and Yang, G., 2016. Deep learning for human activity recognition: A resource efficient implementation on low-power devices. 2016 IEEE 13th International Conference on Wearable and Implantable Body Sensor Networks (BSN),.

Vadapalli, P., 2020. Top 10 Deep Learning Techniques You Should Know About | upGrad blog. [online] upGrad blog. Available at: [Accessed 19 October 2021].

Vrigkas, M., Nikou, C. and Kakadiaris, I., 2015. A Review of Human Activity Recognition Methods. Frontiers in Robotics and AI, 2.

Take a deeper dive into Gender Disparities in Management Positions with our additional resources.