Recruiting AI to Guard Our Home

 Upgrading Technologies change living standards, living standards change daily life problems which need equally upgraded technically advanced solutions. With the increase in advancements, smarter and efficient solutions have come up to promote the safety and transparency of services. 

                  One such field is surveillance. Surveillance includes carefully watching somebody whose activity is fishy. Keeping a watch on suspicious people includes identifying unknown/blacklisted faces and raising alerts when needed. Any such encounter will be read from a video surveillance camera, CCTVs employed will send it to the software/app scanning the footage. Artificial Intelligence enhanced Deep Learning Models will scan this footage using Neural Networks, which will be fed with data of familiar faces and of blacklisted people(in certain cases). After an alert, the concerned person or operator may contact the police or concerned authorities. 

 Taking the technical point of view of the same, a deep learning model will be deployed for scanning of the live video being recorded through the CCTV Camera installed. The pre-Trained Model would find faces and check if it belongs to its Database. This will be done through Neural Networks. Neural Networks in Artificial Intelligence are inspired by the Human Neural system and follow a similar way of remembering faces.

 Suppose you have your memory of having person X’s face and then one day you meet him accidentally. You match the face before your eyes with one in your memory. When your brain finds similarities in the two, it confirms that the person standing in front is person X and then you greet the person to begin the conversation. 

These faces are not saved just as an image but as a calculation of features, features of the face like face shape, eyes, identification marks(not biometrics necessarily) are saved and compared to the input face. This is the place where Neural Networks are developed. A number of images of the same person in different positions, of different sizes(face size incomplete image), different clarity is given to the model. As provided for training as all images are of the same person so the face will have common features which will be recorded. Ones that are uncommon, like some cut, style, mask are not recorded which makes the model get more clarity on the face of the person. 

                 In the practical functioning of the system, CCTV records frames and sends them for face extraction where possible faces present in the frames are saved and tested against prototypes that were used for training. A certain percentage of similarity is calculated as is matched with the value given in settings provided by the user. If the model finds the similarity percentage to be more than the value specified, a positive response is sent to the system controlling alert system. The alert system records the time of entry and sends a signal to the lock and the display/monitor inside the house to make the person (attendant) aware there is someone at the door and it’s safe to let him/her in.

The system finds its application in Homes, Schools, Offices, Banks and various other places where the ’protected/target area’ is accessible to fewer people and proper checks can be made. For example, Homes are mostly accessible to 4-5 members of the family and 1 house service, schools have many highly secure areas like primary sections or transportation facilities (for drivers/family members to take students back) which need high-security to stop all form of frauds and child abuse by strangers, Offices also have departments with confidential details and Banks have treasury and Lockers. These places allow mostly less than 8 people to cross the security checkpoint, which also makes the system more efficient in identifying faces.

       In addition to places stated above, it can be used in places with less number of and where clear video can be recorded & be sent to the device for later procedure. 

 The model currently covers domains with less number of users, in the future, this can be extended to domains where a large number of users are involved. This at times may also fall in the category where the training dataset is very small. This may be some big meeting or conference, marriage(VIP/VVIPs where security needs to be very tight and attendees are too less), and even in places like a prison. Breaching security will be very difficult and alerts will be raised very quickly. With improving efficiency and making the system secure from cyber attacks it can be even installed for high-security areas (Eg: check posts) which will be the final destination for the model.

KUSHAGRA SRIVASTAVA (2000290100085)

B.Tech CSE-3rd-B

KIET GROUP OF INSTITUTIONS, GHAZIABAD