Anamolous Activity Recognition using Deep Learning Techniques

Project Title

Anamolous Activity Recognition using Deep Learning Techniques

Project Area of Specialization Artificial IntelligenceProject Summary

Today we are living in a small digital era where technologies like AI and IOT vows to disturb our lives for the improvement of our way of life. So, we should exploit these headways to practically implement them in our lives. We propose a novel and unique prototype for early detection of traffic accidents and violent incidents. We define violence incidents like brawl, fighting etc. Our prototype working is basically divided into two steps.

Firstly, the prototype will detect any traffic accident or violence incident which will be caught in the eye of the camera. Through image and video processing, if any brawl will happen, camera will detect it as violence incident. Moreover, traffic accident is defined as collision of two objects. Like, two cars colliding, car hitting a person or car colliding with any object like road blockers, footpath, street light poles, etc.

Secondly, upon detecting the incident, prototype will send an alert call to the concerned authorities like patrolling police or nearby police station or hospital so that incident could be encountered on the spot without any delay. Curbing of delay in reporting is our prime initiative in the developing of this project.

Project Objectives

Final prototype will be capable of:

• Robust accident detection
• Detecting violence incidents like brawl, fighting, etc.
• Effective security management system and information system
• Rapid response system will help in early reporting of the incident that will lead to quick response by authorities against the incident

Project Implementation Method

We divided our project into milestones.

• Literature Review:

Before formally kicking off with the project, we firstly did literature review and studied the past work done in this field. We went through different literature review papers and briefly studied the techniques adopted to accomplish the project.

• Data Gathering and Preprocessing

                The next step is data gathering and its preprocessing. We have gathered and compiled the dataset according to our project’s needs. We have divided this data into three categories: violence, traffic accidents, and neutral. Neutral is defined as normal scenario.

• Model selection, training and testing:

After studying literature review and image segmentation, we went for transfer learning approach and choose CNN models. We shortlisted Inception V3 and Yolo V3 and started training our dataset on these models. We defined three scenarios, traffic accident, violence and neutral. Neutral is the scenario where neither traffic nor violence is detected. The next part of our project is to further modify our model for better accuracy. This is done by modification of layers in our model. Then we have started gathering video data to train our model on videos. The main part of our project is the real-time testing of the model, which take the real-time video-feed and convert it into frames and use our model to detect anomalous activity.

• Response management system:

              After prototype will be capable of detecting traffic and violent incidents, we will   generate a response by designing a response management system. Response management system will robustly alert the concerned authorities about incident took place at the respective area by sending an SMS, so that the incident can be quickly encountered. Normally, delay in reporting of the incident leads to loss of lives, money or other assets of people. Our response management system will be robust, efficient and effective solution to all these problems as it will reduce the communication gap to report the incident.

• Code deployment:

Code deployment will be integrating the code into camera and using it practically. This will be part of hardware portion of our project.

Benefits of the Project

Unfortunately, from last couple of years, cases of child abuse and traffic accidents had surged at a massive rate. Still a lot of cases are left unreported. Reducing them is the need of the time. Mostly, delay in reporting of traffic accidents leads to late arriving of ambulance or help from concerned authorities. This delay hangs the life of the patient and can cause loss of life. Similarly, delay in violence cases especially in deserted areas leads to the loss of people’s life, money and assets.

Approximately, 13478 calls were reported in terms of violence cases in Pakistan in first five months of last year. Similarly, 8885 traffic accident cases were reported in Pakistan in 2019. These are the statistics of Pakistan Bureau of Statistics (PBS). Still, most of the cases are left unreported. Using technological advancements, we should look forward towards fulfilling the gap and reporting these incidents timely to the authorities as delay in reporting is the main cause of a huge loss. Pakistan is thirsty of technological advancements. Authorities need these advancements to counter the spread of traffic and violent cases. Rate of surging is alarming and technology should be utilized now to tackle it. Traditional methods are outdated and practicing the new methods is what can possibly be the solution to all these problems. Our project aims to challenge these challenges and with real-time response management system, prototype will quickly alert the concerned authorities to tackle the situation.

 

Technical Details of Final Deliverable

Our project can be bifurcated into three different sections, first one including the training of a deep learning CNN model with appropriate accuracy for identification of road accidents and violence Secondly, upon detecting the incident, prototype will send an alert call to the concerned authorities and the last one embracing the real time implementation. We have already covered more than 50 percent of the milestones by literature review of research papers, detailed scrutiny of prior studies regarding this domain of interest, articulating basic logic and orchestration of coding for them, reanalyzing the logics based on accuracy and precision by varying dependent parameters and in the end finalizing the best possible approach.

1. CCTV IP Camera

• Storage: Up to 128GB
• IR distance: Up to 25m
• Maximum image resolution: 1920x1080 pixels for 2MP, 2592×1520 pixels for 3MP and 2560×1920 pixels for 5MP
• Video bit rate 16Kbps~16Mbps
• Network protocols supported: IPv4/IPv6, TCP, UDP, RTP, RTSP, RTCP, HTTP, HTTPS, DNS, DDNS, DHCP, FTP, NTP, SMTP, SNMP, UPnP, SIP, PPPoE, VLAN, 802.1x
• Power consumption: 2.5W maximum

1. SIM 800L

For response management system, we are using sim 800L module. A SIM card is inserted into SIM 800L and text messages can be generated by pairing it with a microcontroller.

• Working Voltage: 3.4V~4.4V
• Current rating: Consumes approximately <0.7mA of current
• Frequency bands: Quad band supported

1. NVIDIA Jetson Nano

• GPU: 128-core NVIDIA Maxwell architecture-based GPU
• CPU: Quad-core ARM A57
• Video: 4K @ 30 fps and 4K @ 60 fps encode and decode
• Camera: MIPI CSI-2 DPHY lanes, 12 module and 1 developer kit
• Memory: 4GB 64-bit LPDDR4
• Connectivity: Gigabit Ethernet
• Module Size: 70mm x 45mm
• Developer Kit size: 100mm x 80mm

1. GSM Module:

• Input Voltage: 5V~30V
• Current: 8mA in ideal mode and 150mA in communication mode
• Temperature range: -300C-850C

Final Deliverable of the Project Hardware SystemCore Industry EducationOther Industries IT Core Technology Artificial Intelligence(AI)Other Technologies Internet of Things (IoT)Sustainable Development Goals Good Health and Well-Being for PeopleRequired Resources

Item Name	Type	No. of Units	Per Unit Cost (in Rs)	Total (in Rs)
			Total in (Rs)	61500
GSM Module	Equipment	1	2000	2000
NVIDIA Jetson Nano	Equipment	1	35000	35000
SIM 800L	Equipment	1	2000	2000
CCTV IP Camera	Equipment	4	5000	20000
Arduino Mega 2560	Equipment	1	2500	2500