Background. This research centers on tackling the serious global problem of trafficaccidents. With more than a million deaths each year and numerous injuries, it’svital to predict and prevent these accidents. By combining satellite images and dataon accidents, this study uses a mix of advanced learning methods to build a modelthat can foresee accidents. This model aims to improve how accurately we predictaccidents and understand what causes them. Ultimately, this could lead to betterroad safety, smoother maintenance, and even benefits for self-driving cars and insurance. Objective.The objective of this thesis is to create a predictive model that improvesthe accuracy of traffic accident severity forecasts by integrating satellite imagery andhistorical accident data and comparing this model with stand-alone data models.Through this hybrid approach, the aim is to enhance prediction precision and gaindeeper insights into the underlying factors contributing to accidents, thereby potentially aiding in the reduction of accidents and their resulting impact. Method.The proposed method involves doing a literature review to find currentimage recognition models and then experimentation by training a Logistic Regression, Random Forest, SVM classifier, VGG19, and the hybrid model using the CNNand VGG19 and then comparing their performance using metrics mentioned in thethesis work. Results.The performance of the proposed method is evaluated using various metrics, including precision, recall, F1 score, and confusion matrix, on a large datasetof labeled images. The results indicate that a high accuracy of 81.7% is achieved indetecting traffic accident severity through our proposed approach where the modelbuilt on individual structural data and image data got an accuracy of 58.4% and72.5%. The potential utilization of our proposed method can detect safe and dangerous locations for accidents. Conclusion.The predictive modeling of Traffic accidents are performed using thethree different types of datasets which are structural data, satellite images, and acombination of both. The finalized architectures are an SVM classifier, VGG19, anda hybrid input model using CNN and VGG19. These models are compared in orderto find the best-performing approach. The results indicate that our hybrid modelhas the best accuracy with 81.7% indicating a strong performance by the model.
Identifer | oai:union.ndltd.org:UPSALLA1/oai:DiVA.org:bth-25519 |
Date | January 2023 |
Creators | Sandaka, Gowtham Kumar, Madhamsetty, Praveen Kumar |
Publisher | Blekinge Tekniska Högskola, Institutionen för datavetenskap |
Source Sets | DiVA Archive at Upsalla University |
Language | English |
Detected Language | English |
Type | Student thesis, info:eu-repo/semantics/bachelorThesis, text |
Format | application/pdf |
Rights | info:eu-repo/semantics/openAccess |
Page generated in 0.0014 seconds