Expanding vision based ADAS for non structured environments

Abstract

Advanced driver assistance systems (ADAS) become an integral part of almost all modern automotive systems. ADAS have been evolving over a decade and the expansion of vision-based ADAS is quite rapid mainly due to the recent advancements in camera technologies. Most of the vision-based ADAS applications have been developed focusing on structured environment parameters and being tested adequately for those environments whereas they cannot be applied with their current framework as such for non-structured environments due to various limitations. This study presents a comprehensive overview of challenges in expanding the vision-based ADAS for non-structured environments. The major challenges are accurate object detection with deep learning networks, Vulnerable Road User(VRU) path prediction and surrounding vehicle motion prediction. VRU trajectories are very dynamic and extremely challenging for prediction. VRUs tend to dynamically change their course of movement quite often especially in urban scenarios when moving along with heterogeneous traffic elements. Path prediction algorithms help ADAS control systems to keep a watch on those VRUs that are likely to enter into critical zones from non-critical zones. Accurate prediction in advance helps ADAS systems to handle the dynamic behavior effectively without subtle changes in the dynamics and thereby provide better ride handling experience to the occupants. In addition to VRU prediction, predicting road agents intentions holds paramount importance for Autonomous vehicles, especially considering the forthcoming coexistence of ADAS systems with heterogeneous road entities within urban roadways. This study also captures the surrounding vehicles intent to utilize tight lateral spaces in non-lane based environments through the variations of lateral descriptor values. The proposed work included a segmentation detection method for pedestrians and cyclists in a non-structured road environment to improve the accuracy of the popular deep learning networks followed by upper bo