The Deep Learning Based Driver Assistance System was developed with the aim of reducing car accidents caused by driver error, using a deep learning-based object detection algorithm. During our investigation, we discovered that the YoloV3 and YoloV4 algorithms, along with their pretrained Darknet53 and CSP-Darknet53 models, were susceptible to low-light images when detecting objects. To address this issue, we fine-tuned the pre-trained networks using the “Oxford RobotCar Dataset” to minimize the network’s bias towards daytime images in the object detection task. With this issue resolved, we created a pipeline that detects road lanes, calculates the probability of collisions, and provides feedback to the driver.
