Camera Blockage Detection in Autonomous Driving using Deep Neural Networks
N R Vikas, Gaurav Pahwa, Suman Mohanty · 2022 Second International Conference on Computer Science, Engineering and Applications (ICCSEA) · 2022
The autonomous driving industry is evolving at a pace unimaginable a decade ago. Machine Learning (ML) algorithms powered by availability of large quantity of data and specialized processors in form of graphical processing units (GPUs) have accelerated that growth. However, if the data itself is of poor quality, then the machine learning models used to perceive and react to the environment will lack the robustness which is crucial for safety of the Advanced Driver Assistance Systems/ Autonomous Driving (ADAS/AD) applications in next generation vehicles. ADAS/AD applications are fueled by data from sensors like camera, Light Detection and Ranging (LiDAR), Radio Detection and Ranging (RADAR), Inertial Measurement Unit (IMU) and Global Positioning System (GPS). Camera constitutes one of the crucial parts of an AD system which is susceptible to blockage due to environmental factors like rain, dirt, mud etc. Automotive manufacturers conduct data collection campaigns to acquire data to train the ML algorithms. The motivation for this paper comes from data quality assessment which can be performed on data collected during campaigns to identify video segments containing occluded images due to camera blockage. We propose a DNN based on ResNet architecture which utilizes multiple open-source blockage datasets. The model created iteratively tested using augmented data generate using GAN and its learning validated using AI explainability techniques. The result is a robust and generalized model.