Crime forecasting: a machine learning and computer vision approach to crime prediction and preven...
1. In other book, the authors provided insight into how we can achieve crime mapping using satellites. They concluded the following: (1) Visual features contained in satellite imagery can be successfully used as a proxy indicator of crime rates; (2) ConvNets are capable of learning models for crime rate prediction from satellite imagery; (3) Once deep models are used and learned, they can be reused across different cities.
2. Computer vision models combined with machine and deep learning techniques
2.1. A framework for multi-camera video surveillance is presented. The framework is designed so efficiently that it performs all three major activities of a typical police “stake-out”, i.e., detection, representation, and recognition.
2.2. The detection part mixes video streams from multiple cameras to efficiently and reliably extract motion trajectories from videos. The representation helps in concluding the raw trajectory data to construct hierarchical, invariant, and content-rich descriptions of motion events.
2.3. Also, a method is suggested for identifying people for surveillance with the help of a new feature called soft biometry, which includes a person’s height, built, skin tone, shirt and trouser color, motion pattern, and trajectory history to identify and track passengers, which further helps in predicting crime activities.
2.4. Another discussed topic was automated visual surveillance in a realistic scenario and used Knight, which is a multiple camera surveillance and monitoring system. Their major targets were to analyze the detection, tracking, and classification performances. The detection, tracking, and classification accuracies were 97.4%, 96.7%, and 88%, respectively. This again proves the reliability of computer vision models.
2.5. Another experiment was discussed in which they examined the ability of subjects to identify targeted individuals captured by a commercially available video security device. In the first experiment, subjects personally familiar with the targets performed extremely well at identifying them, whereas subjects unfamiliar with the targets performed quite poorly.
2.6. Although these results might not seem to be very conclusive and efficient, police officers with experience in forensic identification performed as poorly as other subjects unfamiliar with the targets. In the second experiment, they asked how familiar subjects could perform so well, and then used the same video device edited clips to obscure the head, body, or gait of the targets.
2.7. Hiding the body or gait produced a small decrease in recognition performance. Hiding the target heads had a dramatic effect on the subject’s ability to recognize the targets. This indicates that even if the quality of the video is low, the head the target was seen and recognized.
2.8. In other of the referenced books, the authors suggested an extremely intriguing research approach in which they claim to prove that looking beyond what is visible is to infer meaning to what is viewed from an image. They even conducted an interesting study on determining where a McDonalds could be located simply from photographs, and provided the possibility of predicting crime.
2.9. This research indicated the presence of some visual cues that are not easily spotted by an average human, but are able to be spotted by a machine, thus enables us to judge whether an area is safe. The authors indicated that numerous factors are often associated with our intuition, which we use to avoid certain areas because they may seem “shady” or “unsafe”.
3. ML techniques used in crime prediction
3.1. A comparative study was carried out between violent crime patterns from the Communities and Crime Unnormalized Dataset versus actual crime statistical data using the open source data mining software Waikato Environment for Knowledge Analysis (WEKA).
3.2. The scope of the project was to prove the efficiency and accuracy of ML algorithms in predicting violent crime patterns and other applications, such as determining criminal hotspots, creating criminal profiles, and learning criminal trends.
3.3. A ML approach is presented for the prediction of crime-related statistics in Philadelphia, United States.
3.4. Algorithms such as logistic regression, KNN, ordinal regression, and tree methods were used to train the datasets to obtain detailed quantitative crime predictions with greater significance.
3.5. Different types of crimes ranging from assaults to cyber fraud were included to match the general pattern of crime in Philadelphia for a particular interval of time. Their algorithm was able to predict whether a crime will occur with an astonishing 69% accuracy, as well as the number of crimes ranging from 1 to 32 with 47% accuracy.
3.6. Also, supervised ML techniques were used to analyze the dataset to carry out data validation, data cleaning, and data visualization on the given dataset.
3.7. The proposed system consists of data collection, data preprocessing, construction of a predictive model, dataset training, dataset testing, and a comparison of algorithms.
3.8. A feature-level data fusion method based on a deep neural network (DNN) is proposed to accurately predict crime occurrence by efficiently fusing multi-model data from several domains with environmental context information.
3.9. The DNN model consists of spatial features, temporal features, and environmental context. The SVM and KDE models had accuracies of 67.01% and 66.33%, respectively, whereas the proposed DNN model had an astonishing accuracy of 84.25%.
3.10. Techniques such as Bayesian neural networks, the Levenberg Marquardt algorithm, and a scaled algorithm were used to analyze and interpret the data, among which the scaled algorithm gave the best result in comparison with the other two techniques. A statistical analysis based on the correlation, analysis of variance, and graphs proved that with the help of the scaled algorithm, the crime rate can be reduced by 78%, implying an accuracy of 0.78.
3.11. A big data and ML technique for behavior analysis and crime prediction is presented. It was concluded that big data is a suitable framework for analyzing crime data because it can provide a high throughput and fault tolerance, analyze extremely large datasets, and generate reliable results, whereas the ML based naïve Bayes algorithm can achieve better predictions using the available datasets.
3.12. The authors considered the development of a crime prediction model using the decision tree (J48) algorithm. When applied in the context of law enforcement and intelligence analysis, J48 holds the promise of mollifying crime rates and is considered the most efficient ML algorithm for the prediction of crime data in the related literature. The J48 classifier was developed using the WEKA tool kit and later trained on a preprocessed crime dataset. The experimental results of the J48 algorithm predicted the unknown category of crime data with an accuracy of 94.25287%.
4. Present technological used in crime detection and prediction
4.1. Currently, there are tools used by police to assist in specific tasks such as listening in on a suspect’s phone call or using a body cam to record some unusual illegal activity.
4.2. One good way of tracking phones is through the use of a stingray, which is a new frontier in police surveillance and can be used to pinpoint a cellphone location by mimicking cellphone towers and broadcasting the signals to trick cellphones within the vicinity to transmit their location and other information.
4.3. A stakeout is the most frequently practiced surveillance technique among police officers and is used to gather information on all types of suspects.
4.4. Another way to conduct surveillance is by using drones, which help in various fields such as mapping cities, chasing suspects, investigating crime scenes and accidents, traffic management and flow, and search and rescue after a disaster.
4.5. Although each of these methods is effective, one point they share in common is that they all work individually, and while the police can use any of these approaches individually or concurrently, having a machine that is able to incorporate the positive aspects of all of these technologies would be highly beneficial.
5. Abstract
5.1. A crime is a deliberate act that can cause physical or psychological harm, as well as property damage or loss, and can lead to punishment by a state or other authority according to the severity of the crime.
5.2. In the current scenario of rapidly increasing crime, traditional crime-solving techniques are unable to deliver results, being slow paced and less efficient.
5.3. The sole purpose of this study is to determine how a combination of ML and computer vision can be used by law agencies or authorities to detect, prevent, and solve crimes at a much more accurate and faster rate.
6. Introduction
6.1. Computer vision is a branch of artificial intelligence that trains the computer to understand and comprehend the visual world, and by doing so, creates a sense of understanding of a machine’s surroundings.
6.2. Research is currently being conducted on the formation of mathematical techniques to recover and make it possible for computers to comprehend 3D images.
6.3. Machine learning (ML) is an application that provides a system with the ability to learn and improve automatically from past experiences without being explicitly programmed.
6.4. Applications of crime forecasting are currently being used by law enforcement in the United States, the United Kingdom, the Netherlands, Germany, and Switzerland.
6.5. The main aim of crime forecasting is to predict crimes before they occur, and thus, the importance of using crime forecasting methods is extremely clear. Furthermore, the prediction of crimes can sometimes be crucial because it may potentially save the life of a victim, prevent lifelong trauma, and avoid damage to private property.
6.6. In this paper, they aim to make an impact by using both ML algorithms and computer vision methods to predict both the nature of a crime and possibly pinpoint a culprit.
6.7. Herein, authors discuss a theory where they combine machine-learning algorithms to act as a database for all recorded crimes in terms of category, along with providing visual knowledge of the surroundings through computer vision techniques, and using the knowledge of such data, they may predict a crime before it occurs.
7. Proposed idea
7.1. After finding and understanding various distinct methods used by the police for surveillance purposes, authors determined the importance of each method. This brought them to the main point of their paper discussion, i.e., their proposed idea, which is to combine the point aspects of Sting Ray, body cams, facial recognition, number plate recognition, and stakeouts.
7.2. The new features are completely computer dependent and hence require human interaction for development; however, once developed, it functions without human interaction and frees humans for other tasks. Let's understand the use of each function.
7.2.1. Core analytics: This includes having knowledge of a variety of statistical techniques, and by using this knowledge, predict future outcomes, which in our case are anything from behavioral instincts to looting a store in the near future.
7.2.2. Neural networks: This is a concept consisting of a large number of algorithms that help in finding the relation between data by acting similar to a human brain, mimicking biological nerve cells and hence trying to think on its own, thus understanding or even predicting a crime scene.
7.2.3. Heuristic engines: These are engines with data regarding antiviruses, and thus knowledge about viruses, increasing the safety of our system as it identifies the type of threat and eliminates it using known antiviruses.
7.2.4. Cryptographic algorithms: Such algorithms are used in two parts. First, they privately encode the known confidential criminal data. Second, they are used to keep the newly discovered potential crime data encrypted.
7.2.5. Recursion processors: These are used to apply the functions of our machine repeatedly to make sure they continuously work and never break the surveillance of the machine.
7.2.6. Bayesian networks: These are probabilistic acyclic graphical models that can be used for a variety of purposes such as prediction, anomaly detection, diagnostics, automated insight, reasoning, time series prediction, and decision making under uncertainty.
7.2.7. Data acquisition: This might be the most important part because our system has to possess the knowledge of previous crimes and learn from them to predict future possible criminal events.
7.2.8. Document processors: These are used after the data collection, primarily for going through, organizing, analyzing, and learning from the data.
7.2.9. Computer linguistics: Using algorithms and learning models, this method is attempting to give a computer the ability to understand human spoken language, which would be ground breaking, allowing a machine to not only identify a human but also understands what the human is saying.
7.2.10. Natural language processor: This is also used by computers to better understand human linguistics.
7.2.11. Voice print identification: This is an interesting application, which tries to distinguish one person’s voice from another, making it even more recognizable and identifiable. It identifies a target with the help of certain characteristics, such as the configuration of the speaker’s mouth and throat, which can be expressed as a mathematical formula.
7.2.12. Gait analysis: This will be used to study human motion, understanding posture while walking. It will be used to better understand the normal pace of a person and thus judge an abnormal pace.
7.2.13. Bio metric identification: This is used to identify individuals by their face, or if possible, identify them by their thumb print stored in few different databases.
7.2.14. Pattern mining: This is a subset of data mining and helps in observing patterns among routine activities. The use of this technology will help us identify if a person is seen an usual number of times behind a pharmacy window at particular time, allowing the machine to alert the authorities.
7.2.15. Intel interpretation: This is also used to make sense of the information gathered, and will include almost all features mentioned above, combining the results of each and making a final meaningful prediction.
7.2.16. Threat detection: A threat will be detected if during the intel processing a certain number of check boxes predefined when making the system are ticked.
7.2.17. Threat classification: As soon as a threat is detected, it is classified, and the threat can then be categorized into criminal case levels, including burglary, murder, or a possible terrorist attack; thus, based on the time line, near or distant future threats might be predictable.