Autonomous vehicles use different sensors to estimate their surroundings, one of which is a stereo camera. However, the efficiency of this sensor in unstructured and heterogeneous traffic has not been studied. This paper discusses and evaluates some state-of-the-art depth estimation algorithms alongside traditional stereo-matching algorithms. The algorithms have been evaluated in various weather conditions and times of the day. The environment contained unstructured and heterogeneous traffic elements such as cyclists, dense traffic, two-wheeled vehicles, and random pedestrians. The paper considers different methods for stereo-matching and generating disparity maps. To ensure that, the results were produced in diverse scenarios, the stereo depth estimation algorithms were evaluated on ApolloScape, and data was generated from CARLA.
Purpose: To study the feasibility of a stereo camera as an obstacle detection sensor for unstructured environments
Design/Methodology/ Approach: Writers studied different algorithms, evaluated them on ApolloScape and generated the data from CARLA.
Findings: From all the tested algorithms, PSMNet has the best performance on average for all values. Unlike other algorithms like noon and sunset, it gives the same results for nightly weather conditions.
Research Limitations/ implications: The writers are going to make the results reproducible and add details to studied algorithms.
Originality/ Value: This is an evaluation and study of Stereo Vision effectiveness in an unstructured environment. This will add to our understanding of computer vision and its application in ADAS and AV.
Keywords: Autonomous Vehicles, Stereo Vision, Traffic Simulation