Deep Learning for Autonomous Vehicle Sensor Data Analysis and Interpretation

Abstract

As the number and the availability of sensors is increasing, deep learning-based data analytics also has already started to be introduced to process and interpret AV sensor data. Features extracted from sensor data are usually high-dimensional, complex, dynamic, noisy, chaotic, redundant, slow-ticking and multi-modal. In order to analyze and process such high-dimensional data, the traditional algorithms failed to cope with the analysis due to their intrinsic limitations. Special feature extraction needs pretty specific pre-knowledge and a lot of experiences by human. But deep learning models can calculate manifold effortly and can drive directly for sensor data and human can interpret only when it is failure case. Investigating modality-separated deep learning models are needed in depth for all sensor datas and to find unsupervised feature extraction methods without label information another expected research area. Training and fine-tuning of deep learning models for big-sized sensor data requires a lot of time and energy, and computer vision GPU and TPUs are important hardware to ease learning processes. Finally, a model optimization, hardware sensors and software architectural research is required for real-time systems [1].