According to America Census Tract data [5], the motorist is the second-highest mode used
to commute to work, as shown in the figure above.
since the adaption of Vision zero, NYC has witnessed several implementation measures
to develop safer intersections such as:
• Increasing the number of bike mileage.
• Increasing the Leading Pedestrian Intervals (LPI).
• Impalement a total of 158 Street Improvement Projects (SIP).
• Installation of the speed camera within the school zone.
• An increase in the number of violations, with a total of 81,609 summonses were
issued to drivers for failure to yield.[4]
This study reviewed the highest number of contributing factors to crashes based on
location. The purpose of the research study is to identify high-risk locations identified by the
contributing factors and the number of collisions resulting in a higher risk of potential collisions.
In order to classify high-risk crash locations, this analysis analyzed the highest number of recurring
crashes, identified by the contributing risk factors and the number of collisions that result in a
location having a higher risk of potential collisions. The analysis uses various statistical models to
generate risk predictions by integrating multiple types of data. The data used historical collisions to
determine high-risk locations; NYC-Open-Data crash data was used as the primary source for
collision data.

Systemic Safety Approach

As outlined in the Federal Highway Administration’s (FHWA) Systemic Safety Project Selection Tool are the following:
The above principles can be described as the methodology to identify high-risk locations for crashes based on a data-driven approach (Preston et al., 2013). These steps assist in finding the type of crashes and the risk factor associated with the collision. The functionality of this model might have the capability to provide some mitigation measures. 

Literature Review

Collision Predictors have been tested in several analyses using similar methods and data frames. The subsequent research used the analogous regression approach to obtain a model and or used the spatial method.
A. Iranitalab, A. Khattak Prediction models for crash severity, allow various agencies to predict the severity of a reported crash with uncertain severity or the severity of crashes that could occur sometime in the future. This paper focuses on the following: comparison of the performance in forecasting traffic crash occurrence of four statistical and machine learning approaches, including Multinomial Logit (MNL), Nearest Neighbor Classification (NNC), Support Vector Machines (SVM), and Random Forests (RF); developing a crash cost-based approach to the comparison of crash severity prediction methods; and exploring the impact on the performance of crash severity prediction models of data clustering methods containing K-means Clustering (KC) and Latent Class Clustering (LCC). Crash data from Nebraska, United States was collected from the 2012-2015 records, and two-vehicle crashes were extracted as the study data. 
The dataset was divided into subsets for training/estimation (2012–2014) and validation (2015). The accurate predictor and proposed approach indicated that NNC had the highest predictive results in both the total and the more severe crashes. RF and SVM had the next two sufficient results, and the worst approach was MNL. Data clustering did not impact SVM's prediction output, but KC improved MNL, NNC, and RF's prediction performance, while LCC improved MNL and RF but decreased NNC's performance. Overall, the accurate prediction rate had almost the exact opposite effects relative to the proposed solution, indicating that neglecting the crash costs could lead to a misjudgment in the choice of the correct prediction process. ⠀
C. Zhang et al. (2014). Focuses on zonal crash prediction and safety assessment are essential in transportation, safety planning, and safety diagnostics. Geographic Information System (GIS) based framework for data integration and safety assessment was introduced. The research established a crash prediction model using a Negative binomial method. The proposed data showed: the crash frequency is correlated with road and traffic characteristics, such as the average free flow rate, the average daily traffic within the zone and the total length of the road within the zone, and social-economic and demographic characteristics, such as the overall population, the percentage of households with high incomes. Three datasets were used for this zonal prediction model: The Traffic Analysis Zones (TAZ) dataset, the traffic and roadway dataset, and the accident datasets. According to the negative binomial model estimation, 
It is possible to estimate the total number of crashes in a TAZ by its Mean expression.
J. Lee et al.  Crash simulations have played a key role in detecting crash hotspots and evaluating safety response initiatives. A variety of macro-level crash prediction models have recently been developed to integrate highway safety considerations into the long-term transport planning process. In this study, the authors defined a series of intersection crash models with macro-level data for seven spatial units for total, severe, pedestrian, and bicycle crashes. The study found that the bicycle crash models performed better in zip-code tabulation area data, and the census-based data-based pedestrian crash models outperform opposing models. It was also identified that intersection crash models could be significantly enhanced for macro-level entities by including random effects. For each crash type, three types of models were developed as follows: Model type (1): crash prediction models with only micro-level variables; Model type (2): micro-level variables and macro-level random-effects crash prediction models; and Model Type (3): crash prediction models of micro-level variables and random-effect macro-level variables. The results showed that Macro-level random-effects only intersection crash forecast models and those with macro-level random effects and variables outperform those with only intersection-level variables. By adding macro-level random effects, the intersection crash prediction models can be improved. With ZIP Code Tabulation Areas (ZCTA) based data, the intersection crash prediction models for total, severe, and bicycle crash models have the highest performance. With Census Tract based data, the intersection crash prediction model for pedestrian crashes performs the best. Finally, pedestrian and bicycle crash simulation findings suggest that multiple macro-level variables may be a robust surrogate exposure for such crashes.
H. Huang et al. (2016) Generally, the level of zonal protection is measured by comparing separate macroscopic variables to aggregated crash statistics on a specific spatial scale. The level of zonal protection was measured by comparing separate macroscopic variables to aggregated crash statistics on a particular spatial scale. The micro-level perspective, in which zonal crashes are determined by summing up all road entities' predicted crashes within the zones of interest. The purpose of this study was to compare these two forms of models for zonal crash prediction. The Bayesian macro-level spatial model with a conditional auto-regressive prior and the Bayesian micro-level spatial joint model was developed and empirically tested. The Optimized Hot Zone Detection method suggested using the benefits of separate macro and micro screening results. The study focused on a three-year data collection for the urban road network. Results have demonstrated that the micro-level model has better overall performance and predictive efficiency the shows an insight into the micro-factors that directly correlate to the severity of an injury and contributes to more direct counter steps. Whereas the macro-level crash review has the benefit of requiring fewer comprehensive details, offering additional input on non-traffic infrastructure problems, and being an essential method for integrating safety concerns into long-term transportation planning. 
M. Abdel-Aty et al. (2013). research with the Geographic Information System (GIS) can analyze crashes with different geographical units. Macro-level safety research is at the center of Transport Safety Planning (TSP) and is crucial in many aspects of safety investment strategy and decision-making. The choice of the spatial unit may vary depending on the variable of the model. In this study, three different crash models for Traffic Analysis Zones (TAZs), Block Groups (BGs), and Census Tracts (CTs) were investigated. Models were developed for total crashes, severe crashes, and pedestrian crashes in this region. The research's primary objective was to analyze and examine the impact of zone heterogeneity (scale and zoning) on these particular types of crash models. These models have been developed based on different road characteristics and census variables. The importance of the explanatory variables was found to be inconsistent between models focused on different zoning systems. Although the variation in variable significance across geographic units has been established, the findings also indicate that the coefficients are rational and explainable across all models.
This analysis's significant results are the coefficients are compatible if these variables are essential in models with the same response variables, even if the geographic units are different. The number of significant variables is influenced by response variables and also by geographic units.
J. Lee et al. (2014) Many crash modeling experiments have concentrated only on the areas where the crash occurred. This study focused on the residence features correlated with the origin of the drivers triggering a traffic accident, the so-called at-fault drivers. Intuitively, it is rational to conclude that the number of at-fault drivers is related to the at-fault driver's residence location's socio-demographic characteristics. Therefore, the key purpose of this analysis is to establish the relationship between the number of at-fault drivers and the zone characteristics of the residence from which the at-fault drivers originated. The Bayesian Poisson-lognormal model was introduced to assess the residential zones' contributing factors to the number of accidents based on faulty drivers.
This research indicates that the crash's occurrence is impacted by road/traffic causes and by many demographic and socio-economic factors of the residential zones. In order to evaluate the relationship between the zonal characteristics and the number of at-fault drivers, the Bayesian Poisson lognormal model was used in this analysis, and the result showed that the exposure measures as 'Log of population,' and 'Proportion of commuters using non-motorized modes' of residence zones were positively associated with the number of at-fault drivers.
On the other hand, 'Proportion of elderly people,' 'Proportion of people working at home,' 'Proportion of commuters whose travel time is less than 15 min' and 'Median Family income' in residential areas have had a negative association with the number of at-fault drivers.
S. Alkheder et al. In this research Analysis, WEKA (Waikato Climate for Knowledge) data-mining software was used to build the artificial neural network (ANN) classifier. The traffic accident data was used in two separate ways to create two classifiers. For testing and validating the first classifier (training set), 90% of the data was used to train the second classifier, and the remaining 10% was used to assess it (testing set). The experimental results showed that the established ANN classifiers could predict accident severity with moderate accuracy. 
The average performance of the training and testing results for model estimation was 81.6 percent and 74.6 %, respectively. Traffic accident data were grouped into three clusters using the k-means algorithm to increase the ANN classifier's estimation accuracy. The results after clustering demonstrated a substantial change in the ANN classifier's prediction accuracy, especially for the training data collection. 
In this work, to verify the efficiency of the ANN model, the ordered probit model was also used as a benchmark. The dependent variable (i.e., degree of injury) has been modified from ordinal to numerical (1, 2, 3, 4) for (minor, moderate, severe, death). The R tool has been used to execute the ordered probit. The ordered probit model revealed how likely this accident would occur in each type of accident (minor, moderate, severe, death). The precision of 59.5 percent was achieved from the probit model, less than the ANN's precision of 74.6 percent.
Z. Elassad et al. The objective of this research is to develop and validate an ensemble fusion structure based on various base classifiers running on fused features and a Meta classifier that learns from the base classifier results to achieve more effective crash predictions. A data-driven methodology was introduced to examine the ability to the convergence of four real-time and continuous categories of features, namely physiological signals, driver maneuvering inputs, vehicle kinematics, and weather covariates, to systematically classify the most robust precursors of the accident using feature selection techniques. Also, a resampling-based scheme, including Bagging and Boosting, is being applied to produce variety in learner combinations comprising Bayesian Learners (BL), k-Nearest Neighbors (kNN), Vector Machine Support (SVM), and Multilayer Perceptron (MLP). In order to ensure that the proposed system provides efficient and stable choices, an imbalance-learning approach was introduced using the Synthetic Minority Oversampling Technique (SMOTE) to solve the issue of class imbalance as a crash case. The results reveal that Boosting has shown the best efficiency within the merger scheme and can reach a maximum of 93.66 percent F1 score and 94.81 percent G-mean with Naïve Bayes, Bayesian Networks, K-NN, and SVM with MLP as a Meta-Classifier. Overall, the methods and results offer new insights into accident detection and can be harnessed as a promising technique to improve intervention efforts relevant to intelligent transport traffic systems.
C. Chen et al. This Research examines driver injury severities in rear-end crashes in New Mexico. Rear-end collisions are a major type of traffic collision in the U.S. A thorough analysis of the function that results in casualties and fatalities is of practical necessity. Decision Table (DT) and Naïve Bayes (NB) approaches have also been commonly used but independently to address classification problems in several fields, except for road safety studies. The study used a two-year rear-end crash dataset; this paper uses a decision table/Naïve Bayes (DTNB) hybrid classifier to determine the probabilistic attributes and predict driver accident results for rear-end crashes. Test results show that the hybrid classifier performs well, as shown by various efficiency metrics, such as precision, F-measurement, receiver operating characteristic (ROC), and area under the ROC curve (AUC). Fifteen significant attributes are significant in predicting driver injury severity, including weather, lighting conditions, road geometry, and driver behavior. The extracted decision rules indicate that heavy vehicle presence, a comfortable traffic environment, low lighting conditions, two-lane rural roads, damaged vehicle injury, and two-vehicle collisions would increase the likelihood of fatal injuries sustained by drivers.
D. Khera et al. Road traffic is an essential part of life, but repetitive accidents cause serious physical harm and property damage. Road  
Traffic Accidents (RTAs) were identified as a major public health issue. The analysis was intended to examine various taxonomies methodologies using the Road Accident Tool WEKA and TANAGRA. The performance was determined by the Naive Bayes, ID3, and Random Tree algorithms. Comparison of data mining algorithm results based on error rate, computational time, precision value, and accuracy. The concept comparison using the WEKA experimenter showed that Naive Bayes surpasses Random Tree and ID3 algorithm with a 50.7 percent accuracy compared to 45.07 and 25.35 percent accuracy using the TANAGRA model. using TANAGRA Random tree outperforms Naive Bayes and ID3 algorithms with an accuracy of 92.95%, 67.6%, and 57.74% respectively.
ACI, and ÖZDEN et al. The analysis of this paper focused on predicting the severity of the motor vehicle accident injuries in Adana, Turkey, using different statistical models to predict the severity of the accidents. The methods were used in this study are K-Nearest Neighbor (KNN), Naive Bayes, Multilayer Perceptron, Decision Tree (DT), Support Vector Machine (SVM). The analysis uses the Regional Traffic Division's traffic accident reports and the Regional Directorate of Meteorology's weather data during 2005-2015; the collision data was recorded in severity as fatal or not fatal. This study's main goal was to determine if the weather condition impacts the collision's severity and any factors for traffic accidents. 
The results indicate that DTC and KNN algorithms yielded slightly more accurate results in classifying fatal instances in both datasets. The accuracy for the KNN was 90.3%, followed by the DTC, which was 90.2%, SVM was recorded at 88.4, MLP was recorded at 87.2%, and the LR was the lowest 82.4%. The analysis of the prediction model's importance measured the Area Under Curve based on the input ranking. The temperature variables were higher based on the methods, identifying that the Maximum Temperature and weather parameters negatively affected all models' classification performance.
Theofilatos, Chen, and Antoniou et al. The main purpose of this study was to investigate the influence of real-time traffic and weather parameters on the frequency of crashes on freeways; no studies are comparing the prediction efficiency of machine learning (ML) and deep learning (DL) models to the best of the results knowledge. The present study contributes to existing information by comparing and validating ML and DL approaches to forecast real-time crash events. Real-time traffic and weather data from Greece's Attica Tollway were connected to historical crash data. The comprehensive data set was divided into subsets of training/estimation (75%) and validation (25%), then structured. Initially, using the training data collection, the ML and DL prediction models were trained/estimated. The models were subsequently compared on the test set based on their performance parameters (accuracy, sensitivity, precision, area under the curve, or AUC). K-nearest neighbor, Naive Bayes, decision tree, random forest, support vector machine, external neural network, and, ultimately, deep neural network were the models considered. Overall, though it outperformed all other candidate variants, the DL model appears to be more fitting. More precisely, relative to other models, the DL model managed to achieve a balanced efficiency among all parameters (total precision = 68.95 percent, sensitivity = 0.521, accuracy = 0.77, AUC = 0.641). However, it is unexpected that, although being much less complex than other models, the Naive Bayes model achieved good efficiency. The analysis results are beneficial since they offer an initial insight into ML and DL models' success.
Hossain and Muromachi et al. Due to recent advances in information systems and traffic sensor technology, the idea of measuring the crash risk for a concise time window in the near future is gaining more practicality. Although some real-time crash prediction models have been suggested, they are still primitive and require substantial real-life improvements. This manuscript investigates the current frameworks' main limitations and proposes alternatives with an updated structure and simulation system to solve them. It uses a random multinomial log model to define the most relevant predictors and the most appropriate detector positions to acquire data to construct such a model. The model was developed using high-resolution detector data obtained from Shibuya 3 and Shin-juku 4 expressways under the authority of Tokyo Metropolitan Expressway Corporation Limited, Japan. The model was then added to the Bayesian belief net (BBN) to create the real-time crash prediction model. It was explicitly designed for the freeway's simple stretches and estimates the probability of establishing a dangerous traffic situation for a particular 250-meter-long road portion within the next 4-9 minutes. The performance appraisal findings reveal that the model can accurately classify 66 percent of possible accidents with a false alarm rate of less than 20 percent at an average threshold value. The findings reveal that the model will correctly identify 66 percent of the accidents with a false alarm rate of less than 20 percent at a threshold value of 4.56 percent. If the threshold value is increased to 7%, the model will still forecast 58% collisions and 87% normal traffic conditions with 82% average classification precision. Moreover, in the event of a threshold value as high as 14%, with just less than 3% false warning, the model classifies 30 percent of the crash cases.
Alajali, Zhou, and Wang The study aims to introduce an accurate intersection traffic prediction by including additional data sources other than road traffic volume data into the prediction model using various decision trees. The analysis also benefits two sets of data collected from the reports of road accidents and roadworks happening near the intersections. All of the data are driven from the Victorian Government Data Directory maintained by the State of Victoria in Australia. The first dataset is the intersection traffic volume. Additionally, this study uses the sensors installed at the traffic signal, reflecting the intersections' real-time traffic volume. They are focusing on the Central Business District (CBD). The second data source was the accident data set, which consists of several attributes, such as accident ID, location coordinates, road number, date, time, and an accident type. The data set also includes the condition of the road surface. As noted earlier, this study aims at the prediction of traffic, particularly at intersections. Addressing several traffic- issues associated with traffic predictions. Pointing out traffic forecasting is a complex and nonlinear problem. One of the significant problems this study will address is the Recurrent patterns vs. Non-recurring events in traffic predictions. Therefore, the implementation of -Spatiotemporal real-time information related to non-recurring events can help predict accurate traffic. A distinctive characteristic of traffic flow is the existence over time of repeated trends, which are useful for traffic prediction models. These events require a nonlinear model, along with the patterns of time and space variations. Another element that this study points out is scalability, which is a crucial traffic prediction requirement since it relies on a large amount of historical data and real-time data sources. The study provides a variety of factors in terms of related work. The traffic modeling for this analysis only evaluates the short-term traffic prediction can be divided into two major segments, and they are Parametric methods and Nonparametric methods. The parametric methods: it is a flexible family of models and estimates the model parameters based on the training data; the paper's abstract gives the topic a solid, concise sense by explicitly outlining the approaches used to solve the proposed problems. However, to make the introduction more concrete, the authors may wish to make multiple references to support the argument made in identifying the methods of the decision trees and how they are going to be deployed to address the presented problems in regards to intersection traffic predictions. The authors might want to include another statement with descriptions of some of this technology's implementations, along with relevant sources. In the related work, the study identifies the methodologies in which will be used to address the intersection traffic predictions. As was pointed out earlier, the study depends on two modeling methods. They are parametric methods and Nonparametric methods. Both of these models provide numerous benefits, and they are as follows. The parametric methods' benefits are as follows: It is a flexible family of models and estimates the model parameters based on the training data. Auto-Regressive Integrated Moving Average (ARIMA), introducing a space-time ARIMA model for urban traffic forecasting. This introduces the proposal of Support Vector Regression (SVR), used for traffic speed prediction. SVR is one of the methods that achieve good accurate results. On the other hand, the study points out the benefits of Nonparametric methods, and they are the following: It is widely adopted for predicting traffic flow, as they are considered to be more suitable than other methods for capturing the nonlinear and stochastic nature of traffic flow. Also, this method uses the k-Nearest Neighbors (KNN) for one stop-ahead traffic prediction. With supporting, methods of Radial Basis Function (RBF) and Artificial Bee Colony (ABC), considered the nonlinear correlation between spatial and temporal features. Also, with a proposed Bayesian Multivariate adaptive Regression (MAR) method for accurate and interpretable traffic prediction. Fast Incremental Model Trees with Drift Detection (FIMT-DD) algorithm. They highlight the use of the FIMT-DD traffic analysis and visualization process. This work expands the FIMT-DD approach by incorporating incidents and roadwork data with regular traffic volume data to forecast the correct traffic at intersections. This paper's objective gives a clear understanding of an approach. However, related work does not identify the purpose of using both methods. Which is assumes that the use of the parametric methods to calculate the errors in the dataset. Also, it is assuming that the nonparametric method was used to calculate the confidence mean. The research equipment is quite standard and appropriate for the research, particularly because the main focus of the paper intersection traffic predictions, possibly, in the related work should have related the use of the decision trees. One important aspect of the research paper underlines some limitations of previous studies related to traffic prediction that considers an online learning approach that focuses on typical and atypical traffic conditions. 
The outlined benefits of using Decision Trees are as follows: the decision Tree will allow decision-makers to manage priority improvements regarding transportation safety measures. Ensemble decision trees were developed to increase decision tree models' performance by combining multiple weak predictors to obtain more accurate predictions.
This research focused on if accidents and roadworks can dramatically influence traffic patterns at intersections; a novel approach to intersection traffic prediction has been suggested, which involves integrating several data sources for model training. Three ensemble decision tree algorithms (GBRT, RF, and XGBoost) have been adopted to train the prediction and model in a batch learning form. It has also introduced an interactive learning system in which the FIMT-DD algorithm was introduced to update the real-time model. The authors may also want to provide a more comprehensive discussion about each decision tree's result and how each would benefit through traffic prediction at the intersections.

Data Used

The data frames used for the study were as follow the five years (2014-2019) of Collision data derived from the New York City Police Department. Extracting High-Risk Locations for Collision Citywide (100 Intersections) and extracting 2019 to generate a prediction model. As a supporting dataset, the population dataset derived from the American Community Survey, 5-year estimate. Next, using the PUMA tabular data extracting the total Population per zip code and using the Public Use Microdata Area (PUMA) shapefile (Cartographic Boundary Shapefiles) New York to ArcMap. Using the Neighborhood Tabulation Areas (NTAs) using the Borough boundary shapefile from NYC OpenData to clip to shoreline ensures that the analysis focused on the City Right of Way. Also, this analysis uses MapPLUTO data from BYTES of the Big Apple for the land use identifications.
Safety Performance Functions (SPFs) explain the statistical associations between crash predictions and the importance of identifying the crash factors. It is important to consider from the literature the causes that increase the level of risk for these types of collisions in order to build a risk-based predicting model. This section began by examining the modeling strategies that have been used to explain roadway collisions.[6] The study then examined the literature on risk factors for roadway collisions, including accident characteristics of roadway design (Pavement condition) associated with increased risk.

Methodology

Python

The datasets were arranged by merging three different Land use tables, Population per zip code, Vision Zero Priority Intersections, and Collision data. For land-use, the main focus was to review the demographic data based on the crashes' location. The analytical tool to organize and simplify the data will be Python (package: Pandas), to understand the need for vision zero to include street geometry/ location as a collision factor. Using Python Data cleaning with Pandas and NumPy using the following functions:
Python-Pandas was used to identify the locations with the highest number of collisions.