Deep Learning based Automated Braking Decision-making for Advanced Driver Assistance System

Advanced Driver Assistance System(ADAS) is a technology that enhance vehicle safety and driving experience and provide drivers the assistance needed for efficient decision making. With the advancement of artificial intelligence technologies in ADAS, reducing traffic accidents remains a global priority to enhance vehicle safety. Traditionally, automakers have relied on a mix of sensors for intelligent braking systems. This paper presents a method that focuses solely on image datasets to develop an Automated Braking System (ABS) eliminating the need for expensive sensors like LiDAR. Just as humans can drive using their eyes, cars should also rely solely on cameras. Utilizing the Berkeley Deep Drive (BDD100K) image dataset, our methodology begins with preprocessing, object detection and then YOLOv8 based bounding box annotation. In the next stage, a pre-trained CNN, ResNet18, is trained with the annotated dataset to train the model that predicts the braking decisions based on calculated object distances from images. Experimental Results obtained a greater accuracy of 94.4%. Experimental analysis for precision, recall, and F1-score are strong for both ‘brake’ and ‘no brake’ decisions, with scores of 0.95, 0.93, and 0.94 for ‘brake’ and 0.92, 0.93, and 0.92 for ‘no brake’, respectively. As Elon Musk famously stated, “Relying on LiDAR is a dead-end. It is an expensive, over-the-top solution for something that does not need it.” This methodology highlights the potential of a purely image-based approach, aligning with Musk’s vision of cost-effective and efficient ADAS solutions. By leveraging deep learning and image processing, this work offers a viable alternative to traditional sensor-reliant systems, potentially transforming the future of automated braking technology.