Explainable Deep Learning-Based Driving Decision Prediction for Autonomous Vehicles Using Grad-CAM
Explainable Deep Learning-Based Driving Decision Prediction for Autonomous Vehicles Using Grad-CAM
Authors:
Sudesh Sanjay Kadukar
Department of E&TC Engineering
D.Y. Patil Education Society School of Engineering and Management
Kolhapur, India kadukarsudesh18@gmail.com
Sushant Sanjay Zende
Department of E&TC Engineering
D.Y. Patil Education Society School of Engineering and Management
Kolhapur, India sushantzende014@gmail.com
Prasad Ashok Chimane
Department of E&TC Engineering
D.Y. Patil Education Society School of Engineering and Management
Kolhapur, India prasadofficial@gmail.com
Shreyash Arjun Patil
Department of Computer Science and Engineering
D.Y. Patil Education Society School of Engineering and Management
Kolhapur, India shreyashpatil2406@gmail.com
Abstract—Autonomous cars employ deep learning models to interpret the road scene and generate driving commands such as turning, braking, accelerating, and going straight. Convolutional neural networks can achieve high prediction accuracy, but their internal decision-making process is often hard to interpret. This black-box behaviour poses problems in safety validation, failure analysis, model debugging and passenger trust. This paper presents an explainable deep learning framework for autonomous driving decision prediction from front-facing road images. A CARLA-based simulated driving dataset is used to collect road images along with steering, throttle, brake, speed and traffic information. The continuous vehicle-control values are translated into five driving-action classes, i.e., move forward, turn left, turn right, apply brake, and accelerate. We train and compare a simple Convolutional Neural Network, the MobileNetV2 and the ResNet50. To identify the image regions that contributed to each prediction, we use the Gradient-weighted Class Activation Mapping, named Grad-CAM. The generated heatmaps help to identify whether the model is attending to relevant objects like pedestrians, lane markings, nearby vehicles, traffic signals, and obstacles. The models are assessed in terms of accuracy, precision, recall, F1-score, confusion matrix, inference time, and explanation relevance. The framework we propose targets the improvement of the transparency, reliability and human understandability of decisions in autonomous driving.
Index Terms—autonomous vehicles, explainable artificial in-telligence, deep learning, Grad-CAM, MobileNetV2, ResNet50, CARLA, predicting driving decisions