Detection and Analysis of Crime using AI

The crime rate is an ever-increasing menace in our society. Especially with the massive urbanization and expanding cities, it has become paramount that Police and Law Enforcement Authorities can get quick information about the happenings of crime. It has been widely observed that the crime happening

Project Title

Detection and Analysis of Crime using AI

Project Area of Specialization

Artificial Intelligence

Project Summary

The crime rate is an ever-increasing menace in our society. Especially with the massive urbanization and expanding cities, it has become paramount that Police and Law Enforcement Authorities can get quick information about the happenings of crime. It has been widely observed that the crime happenings are usually reported after the
crime has taken place and culprits have escaped. Due to fear of the fatality, people do not take any measure which can inform the Police of the crime when it is underway,
even when it is possible to do so.

It is known that almost all of the corporate offices, banks and commercial buildings have already installed CCTV Cameras. Our proposed solution is to use this existing infrastructure and use the feed from these cameras to intelligently detect weapons in the person’s hands. And as soon as the weapon is detected in the camera feed, a response will be triggered which will alert the concerned authorities that crime has started. For this purpose, an embedded device has to be attached to the camera feeds which will be able to process this camera feed. A weapon detection model, similar to the ones discussed in previous section, will also be placed on this embedded device. Videos from
the cameras will pass through this model and whenever the weapon is detected, a response will be generated.
Moreover, we will provide support for multiple cameras to be processed by a single embedded device. This will have many advantages. First of all, the cost will be reduced to a
great extent as single device will be processing many different cameras. Moreover, use of multiple cameras will also make it easier for the system to detect the presence of weapon as camera feed from different angles is taken into consideration.

As soon as crime starts to take place, nearest police station will be immediately informed about the location of the detected weapon. This way the response time of Police can be greatly enhanced.

To detect presence of weapons, an AI model will be build on the embedded device. Camera feed will pass through this AI model and then detected weapons will be marked. 

Project Objectives

Main objective of the project is to improve the safety and security infrastructure. It is very important that crime happenings can be quickly conveyed to relevant Law Enforcement Agencies. Usually the criminals get away with robberies and the incident is only reported after it has taken place and all the culprits have escaped. Then it becomes nearly impossible for Police to catch them. 

Our objective is to develop an intelligent system which can detect the crime taking place by looking at the video feed coming from Live Cameras. Whenever the crime starts to occur, system will automatically detect and send alert to police that crime has started. It is quite possible that police arrives in time to prevent the crime from completely taking palce. 

For this project our objectives can be described as below

• Development of a system which can detect crime from the Camera feed
• Maintain the performance of system in Real time
• Development of mechanism to deliver instant alerts to Police Station. 

Furthermore, this project has a lot of expansion potential. 

Future objectives can be to develop a full security system based on AI. As soon as crime has is detected, Police Drones will depart from police station and reach the crime spot. The data about person associated with crime will be shared with these drones from the internal bank/building cameras. As soon as culprits leave the place, drone will track them, giving police the real time information about the location of suspects. 

Project Implementation Method

Introduction:
In essence, the main part of the project is selecting and then training a proper object detection model which can be successfully used to detect weapons. In order to train the
model and later on, to test the model, a good collection of dataset is required. After training, model has to be run in such synchronization with hardware that model gives
the best possible performance on the hardware.

Model Selection:

Model Selection is an important part of this project. There is a large bundle of models available which can perform object detection very well. Few models which are available
include Faster-RCNN, YOLOv3-Tiny and SSD-Mobilenet-V2. All of them give reasonably good accuracy when it comes to detection of objects from images and videos. In order to make the selection between different models, their structural design and their performance on our particular use case are important parameters to consider. Generally, it is important to keep a track of following points:

• accuracy on near objects
• accuracy on far-away/small objects
• accuracy from different postures and movements
• real-time working speed
• real-time accuracy

Dataset Collection and Annotation:

Next crucial step is the collection of the relevant dataset. As we are aiming to detect pisols in hands, we have searched online to collect the already existing datasets which
are related to our use here. As already existing data is not sufficient enough to train the model to its best accuracy, we are in the process of developing our own dataset. Collection and annotation of dataset is a time consuming task. After the dataset is collected, the annotation of dataset has to be done. In our case, the weapons/pistols present in the images will be annotated and labelled. It is also important, at this stage, to keep under consideration about how the model is performing on current dataset.For
example, which kinds of images the model is working better and which kind of images
is it not classifying correctly. Further efforts to collect dataset must be made in that
direction on which model is not performing well already

Model Training: 

Model is initialized in Tensorflow. The training dataset images are pre-processed and
then fed into the model. Model tries to learn the pistol from the images by the technique
called supervised learning. Ultimate goal is to reduce the difference or error between the
model output and the actual labels in the image. When the error is sufficiently low, the
training is stopped and model is saved.
 

Deployment on Hardware:

Once the model is trained, it is ready to be deployed on the hardware. Hardware we have chosen is Jetson Nano for running the AI Model. This hardware is capable to running small AI models and run them quite efficiently. When model is run on hardware, then input camera is also attached to the Jetson Nano and feed coming from that camera is analysed by the model. This model will detect crime/weapons.

 

Benefits of the Project

Many commercial buildings, banks and corporate offices have already existing security systems. But the issue with these systems is that they are not rapid and
intelligent enough to detect the crime as soon as it starts to take place. For example, many infrastructures have CCTV cameras installed as the part of security system. But
these cameras are only useful in identification of criminals after the crime has taken place.Most of the time, criminals are wearing masks to keep their faces hidden and it is known
that these cameras are not useful enough in finding the criminals.

Almost all the banking institutionhave protective cameras installed for surveillance. Those already existing systems can
be enhanced to automatically detect weapons and report it to nearest Police patrol or Police Station. Day by day, congestion is increasing in the banks. And rate of attempted
robberies in banks and ATMs are increasing. Therefore banks can be a major market for this product. Similarly, the office buildings where security is important are also included
in the application areas of this project.

Moreover, due to urbanization and population rise, the congestion in cities is increasing. The concept of safe cities is beginning to take a definitive shape. Many countries
in the world are installing CCTV cameras throughout the cities to increase surveillance so that safety of citizens can be ensured. It will be great to back those systems with
weapon detection algorithms which can automatically detect weapons and report where the weapon has been detected.

Existing systems also include metal detectors which can give various levels of indications depending upon the type of metal detected. But in most of the scenarios, guards
are very lax and allow people to take mobile phones and other metal devices when they pass through the detectors. Hence it cannot be known for certain if the person is carrying
a weapon or not. In addition to this, they are expensive to install and equally expensive to maintain due to electricity costs. One more issue with these metal detectors is that
they are unable to detect weapons which are made from plastic, following 3D modelling techniques. These plastic weapons are equally lethal and can easily bypass metal scanners, rendering them absolutely useless.
 

This camera based method will provide fool-proof security even in those cases when the weapons are manufactured from plastic using 3D printing. Camera based detection technique will not allow such weapons to escape.
 

Technical Details of Final Deliverable

1. Data Collection and Annotation:

 Good quality can be described in terms of various
parameters:
• Sufficiently large
• Weapons of different sizes
• Different angle to look at weapons
• Good resolution
• Must be similar to target environment
If all of the above stated parameters are met, dataset is considered of sufficiently good
quality.
Annotation in our case referes to adding bounding boxes on the weapons/pistols in the
image. We have used the annotation tool called ’LabelImg’. A label .xml file is generated corresponding to each image. For our use case, we are using the labelling format of
’PASCAL VOC’. This format is widely used and is supported by TensorFlow Object Detection API.
Care has to be taken during the labelling of images. Only the selected region of interest
area has to be labelled. Making large bounding boxes which are covering the background
result in the addition of noise in the data. This will result in reduced accuracy of the
model.
 

2. Google Colaboratory:

Google has provided
free access to GPUs for training of models through CoLab. CoLab provides a notebook interface for a Linux based Virtual Machine. It gives access to 12 GB RAM and 12-15
GB GPU resources, all free of cost. Therefore, it is very helpful to use these free resources
to train our TensorFlow models. In
order to train the model, a massive GPU is required as there are a lot of images which model needs to train on. GPU will offer a large number of parallel computations and
hence the model will be able to train considerably in lesser time

4. Deep Learning Model

The machine learning models which we have selected for this project are SSD-MobilenetV2. SSD-Mobilenet-V2 is the combination of two object detection models: SSD and MobilenetV2.
SSD stands for Single-Shot-Detection. This model calculates the output of one single
pass of the data through the model. Mobilenet-V2 has proven
to be faster than most of the models when it comes to object detection. Its architecture
is similar to most object detection models. On the base, it has a feature extractor model
which produce a single array tensor of features.
 

5. Model Training

TensorFlow Object Detection API is used to train the model. This is the most important
part of the project as training heavily impacts the results of the model.
 

6. Jetson Nano

Jetson Nano is a low-cost, embedded system device with the on board GPU. It is an
ideal low cost device to perform inference using the machine learning model. It has 128
core NVIDIA Maxwell GPU which can run small models like SSD-Mobilenet-V2 very
efficiently.
 

7. TensorRT

TensorRT is the programmable inference accelerator for NVIDIA GPUs. It performs
various optimizations on the model which results in great performance and extremely low
latency. It makes up the middle layer between model and the hardware like GPU. Its
backend is in C++.
 

Final Deliverable of the Project

HW/SW integrated system

Core Industry

Security

Other Industries

Core Technology

Artificial Intelligence(AI)

Other Technologies

Sustainable Development Goals

Industry, Innovation and Infrastructure, Sustainable Cities and Communities

Required Resources

If you need this project, please contact me on contact@adikhanofficial.com