Development of a Deep Learning-Based System for Enhanced Blind Spot Detection and Lane Departure Warning for the Kayoola Buses

Abstract

Deep learning-based advanced driver assistance systems (ADAS) have attracted interest from researchers due to their impact on improving vehicle safety and reducing road traffic accidents. In Uganda, road accidents have continued to soar with an increase of up to 42% in 2021 due to the growing road traffic density. To curb the high rates of road accidents, especially for heavy duty vehicles, Kiira Motors Corporation a state-owned mobility solutions enterprise needs advanced driver assistance systems for improved safety of their market entry products the Kayoola buses. This project presents an approach to vehicular safety enhancement through the implementation of a Lane Departure Warning (LDW) and Blind Spot Detection system (BSD) using advanced deep learning algorithms that will be able to alert the driver using the graphical user interface, and auditory feedback. The system was developed based on the MobileNet architecture and the Kayoola Buses manufactured by Kiira Motors Corporation were used as the project case study. A purposive sampling technique was used to select the study participants focusing on targets automotive manufacturers, bus companies, cargo truck operators, and passengers. Two distinct datasets which included the DET dataset with raw event data from 5424 images of 1280×800 pixels and the TuSimple dataset of 6,408 road images specifically captured on highways were used for model training. The resultant BSD and LDW system are realized on the Raspberry Pi platform, incorporating diverse sensors which include radar sensors, ultrasonic sensors, gyroscope and accelerometer sensors. By combining these advanced features, the study not only bridges an essential research void but also offers a practical resolution to pressing road safety concerns in the East African context. The implementation of a BSD and LDW system through deep learning techniques marks a pivotal advancement in vehicular safety. The lane detection model was tested on Dataset for Lane Extraction (DET) and TuSimple datasets. Our model attained a mean model accuracy (F1 Score) of 77.59% and a mean IoU of 65.26% on the DET and an overall accuracy of 97.96% on the TuSimple dataset. User acceptance tests were carried out to validate and ascertain whether the developed system addressed the needs of the prospective users. The tests were carried out with a total of 150 users to validate the functionality of the system. The anticipated real-world implementation is poised to substantiate the system's effectiveness, thereby contributing to safer roads regionally and inspiring innovation in automotive engineering by leveraging artificial intelligence.