The types of collisions during road crashes and the factors contributing to them are many. Chances of collisions can be reduced by predicting the probability of types of collisions at various locations and providing drivers with real-time details of the risk of collision while they travel. This paper aims to predict the severity of collisions at intersections, specifically two types of collisions—‘Property damage’ and ‘Injury Collision’, through machine learning models. The paper uses a dataset from the Kaggle website which the police department of Seattle regularly updates. It consists of values of different variables such as road condition, weather condition, junction type, and the attention span of the driver, hitting a parked car, light condition, and speeding. Four different conventional machine learning algorithms, namely, Random Forest, K-Nearest Neighbour, Decision tree, and Linear Regression are applied for collision prediction. The Random Forest model is observed to have the highest accuracy (0.7506) followed by the K-Nearest Neighbours algorithm (0.7442), Decision Tree (0.7385) and Linear regression (0.6999). The most featured attributes were weather conditions, road conditions and temperature and predicted more property damage than injury collision. An ensemble of the above models is also tried for crash prediction and observed as with the highest accuracy (0.7519).