Classifying Road Accidents and Forecasting Level of Risk Based on a Combined PCA-LPP and DBSCAN Method

Road traffic accidents are one of the major problems causing large numbers of casualties and property losses worldwide. By classifying road traffic accidents and predicting risk levels, it becomes possible to identify high-risk vehicles and reduce the probability of accidents and casualties. Traffic accidents are often influenced by multiple factors such as environment, weather, road conditions, and infrastructure, but existing accident impact analysis methods lack comprehensive research on traffic accident data. Therefore, this paper proposes a traffic accident classification model that incorporates an improved dimensionality reduction algorithm called PCA-LPP, which measures the similarity between data of different levels to achieve secondary dimensionality reduction. The model utilizes a large-scale traffic accident dataset and applies the DBSCAN algorithm to partition the accident data into risk areas. By training the spatial representations of different risk levels iteratively, the model could assess the risk levels in simulated vehicle environments. Experimental results demonstrate the effectiveness of the proposed approach. Comparative experiments on large-scale traffic data reduced to different dimensions show that the PCA-LPP algorithm achieves higher correlation between the reduced features and sample categories compared to traditional PCA. Moreover, when handling complex and sporadic traffic accident data, the density-based DBSCAN clustering algorithm achieves a purity of 0.942 9, a Rand index of 0.946 2, and a mutual information index of 0.678 4. Comparing these results with traditional algorithms like K-means and spectral clustering, DBSCAN consistently outperforms them in various evaluation metrics. Additionally, visual analysis of the classification results indicates that the proposed model reduces the influence of noisy data. Finally, an ablation experiment confirms that the PCA-LPP algorithm with secondary dimensionality reduction achieves the highest evaluation metrics. The confusion matrix of the prediction results shows that the model achieves precision rates of 85.77%, 70.78%, and 80.65% for different risk levels, further validating its effectiveness and practicality.