Robust visual detection of brake-lights in front for commercialized dashboard camera

PLOS ONE RESEARCH ARTICLE Robust visual detection of brake-lights in front for commercialized dashboard camera Jiyong Moon ID, Seongsik Park ID* Department of Artificial Intelligence, Dongguk University, Seoul, Korea * a1111111111 a1111111111 a1111111111 a1111111111 a1111111111 OPEN ACCESS Citation: Moon J, Park S (2023) Robust visual detection of brake-lights in front for commercialized dashboard camera. PLoS ONE 18(8): e0289700. https://doi.org/10.1371/journal. pone.0289700 Editor: Agbotiname Lucky Imoize, University of Lagos Faculty of Engineering, NIGERIA Received: October 13, 2022 Accepted: July 24, 2023 Published: August 11, 2023 Copyright: © 2023 Moon, Park. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability Statement: All data files are available from Harvard Dataverse Network. Users can find our data at: https://doi.org/10.7910/DVN/ GXFRCQ. Funding: 1. Seongsik Park No.2021R1F1A1062982 Ministry of Science and ICT, MSIT https://www.msit.go.kr/eng/index.do The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. 2. Seongsik Park No.IITP-2021-2020-0-01789 Institute for Information & Communications Technology Abstract The collision avoidance system (CAS) is an essential system for safe driving that alerts the driver or automatically applies the brakes in an expected situation of a vehicle collision. To realize this, an autonomous system that can quickly and precisely detect brake-lights of preceding vehicle is essential and this should works well in various environments for safety reason. Our proposed vision algorithm solves these objectives focusing on simple color features rather than a learning algorithm with a high computational cost, since our target system is a real-time embedded device, i.e., forward-facing dashboard camera. However, the existing feature-based algorithms are vulnerable to the ambient noise (noise problem), and cannot be flexibly applied to various environments (applicability problem). Therefore, our method is divided into two stages: rear-lights region detection using gamma correction for noise problem, and brake-lights detection using HSV color space for applicability problem, respectively. (i) Rear-lights region detection: we confirm the presence of the vehicle in front and derive the rear-lights region, and used non-linear mapping of gamma correction to make the detected region robust to noise. (ii) Brake-lights detection: from the detected rearlights region, we extract color features using the HSV color range so that we can classify brake on and off in various conditions. Experimental results show that our algorithm overcomes the noise problem and applicability problem in various environments. Introduction Whether it is used as a driver assistance system or as a component of autonomous driving, one of the most important factors for safe driving is the collision avoidance system (CAS) [1]. CAS recognizes a collision in advance and avoids it or mitigates the damage from it. Therefore, CAS must recognize the situation around the vehicle and infer information from it. When a collision is expected, the system should either send a warning to the driver or act autonomously without any driver input, i.e., by automatically applying the brakes [2]. In particular, CAS is important in rear-end collision, one of the most frequent traffic accidents [3]. The advanced CAS detects and responds to danger in a rear-end collision situation based on distance information using distance measuring sensor or surrounding context information using vision sensor. Distance measuring sensors, e.g., radars and LiDAR, and their algorithms [4–6], demonstrates significantly high performance; however, it is too expensive to be commercialized [2]. PLOS ONE | https://doi.org/10.1371/journal.pone.0289700 August 11, 2023 1 / 23 PLOS ONE Planning & Evaluation https://www.iitp.kr/en/main. it The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Competing interests: The authors have declared that no competing interests exist. Robust visual detection of brake-lights in front for commercialized dashboard camera But more importantly, the sensor-based method has a disadvantage in rear-end collision; they are operated by estimating the distance to the vehicle in front [7]. Therefore, the distance from the vehicle in front must be sufficiently close or close at a fast speed for the sensor-based CAS to recognize and respond to a rear-end collision situation in advance. This constraint makes it difficult for the sensor-based method to respond flexibly in the rear-end collision situations. Furthermore, there is a possibility of missing the right moment when it should have responded more quickly. On the other hand, vision sensors installed in vehicles and their intelligent methods can see, detect, and distinguish objects directly, owing to state-of-the-art lane detection, vehicle detection, and object identification technologies in computer vision [8]. Especially, vision-based method is much more suitable for the rear-end collision situation than the sensor-based method, because it is independent to distance information; it can acquire high-level information, such as the lane to which the vehicle in front belongs and whether the brake of the vehicle in front is on or off through the optical camera. This high-level information can help predict the braking state of the vehicle in front more accurately and flexibly, allowing the vision-based method to cope with rear-end collision situations more efficiently than the sensor-based method. In accordance with the above, we propose a vision-based algorithm to efficiently cope with rear-end collision situations that can maximize the strength of the vision-based approach. Various vision-based algorithms have been proposed. Some of these methods use machine learning or deep learning algorithms for the detection of rear-lights region and brake-lights [9–11]. These methods perform the detection process by extracting spatial information from the input image. Although these learning algorithms show relatively good performance, they are difficult to use in real-time CAS because of their large computational cost [2], especially in the case of deep learning [1]. Additionally, learning-based methods require considerable training time along with the acquisition/annotation of images/videos to train the model. These burdens may be greater depending on the model size [12]. Therefore, we propose an algorithm based on basic features such as color to operate in real-time and do not require model training (featurebased method). Feature-based methods implement rear-lights region detection and brake-lights detectio (...truncated)