Multi-Sensor Information Fusion for Determining Road Quality for Semi-Autonomous Vehicles 2022-28-0004
Pothole detection in Intelligent Transportation Systems (ITS) vehicles has been a part of Advanced driver-assistance systems (ADAS) for a long time. Various sensors have been used for this purpose so far: Accelerometer, Gyroscope, etc. However, the fusion of multiple modalities of information from different sensors remains a challenge, mainly owing to the different sampling rates and varying frame rates used by each sensor. Other sensor types like Radar and LIDAR, though precise, are difficult to use, thus forcing us to look for low-cost solutions. Our proposed work uses Accelerometer and Gyroscope sensor fusion to predict pothole presence in Indian scenarios. Previous works have mainly dealt with predicting potholes with data collected using either traditional machine learning techniques like Decision trees, (Support Vector Machines (SVM)’s and Light Gradient Boosting Machine (LGBM) and deep learning methods using neural networks and attention mechanisms. In this work, the main focus is on using Convolution Neural Network-based methods to extract information from the sensor data after appropriate preprocessing. Notably, time series models such as Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Transformers integrated with lightweight attention units are then used to classify and predict the presence of potholes based on the information extracted. Further, different classes of neural networks and transformers like Involution Neural networks are investigated for reliable predictions. Subsequently, this information is used to predict potholes and may be used to develop a Road Quality index for Indian Roads which will indicate the quality of a given stretch of road. This paper demonstrates how the proposed method would provide greater accuracy for the prediction of potholes when a vehicle passes through a particular road.