Autonomous vehicles are of great importance in many respects. Autonomous vehicles can provide greater road safety and less accidents by reducing driver errors such as drugged driving, speeding and distraction. Because they require no human interaction, autonomous vehicles can ensure more personal freedom for people with disabilities(such as blind). Autonomous vehicles can provide drivers to spend their time more valuable by both shortening the driving time and enabling them to deal with more productive tasks while driving. Autonomous vehicles can maintain a safe and consistent distance from the vehicles in front of them, helping to reduce the number of stop-and-go waves. This will provide higher traffic efficiency and decrease the time spent on roads for both people and cars. Autonomous vehicles also have the potential to optimize the fuel use and reduce carbon emissions. (can have more numerical data)
Otonom araçlar birçok açıdan büyük öneme taşıyorlar. Otonom araçlar hızlı sürme, dikkat dağınıklığı ve alkollü sürüş gibi sürücüye bağlı olan hataları azaltarak yol güvenliğini büyük ölçüde arttırıp yolda gerçekleşen kazaları azaltabilir. Otonom araçlar insan etkileşimine ihtiyaç olmadan çalışabildikleri için araba sürme engelleri bulunan insanlara daha fazla kişisel özgürlük sağlayabilir. Hem sürüş süresini kısaltarak hem de sürüş sırasında daha verimli işlerle ilgilenmelerini sağlayarak sürücülerin vakitlerini daha değerli geçirmelerini sağlayabilir. Otonom araçlar önlerindeki araçlarla aralarında güvenli ve tutarlı bir mesafeyi sürekli olarak koruyarak trafiğin en büyük nedenlerinden biri olan dur-kalk dalgalarının sayısını azaltabilir. Bu durum trafik sorununu büyük ölçüde çözebilir ve hem insanlar hem de araçlar için yolda geçirilen süreyi azaltabilir. Bunların dışında yakıt tüketimini optimize ederek ve sürüş süresini azaltarak karbon emisyonunu da büyük ölçüde azaltabilir.
Çözüm arayışı , carla (Şafak Erkam)
A sensor fusion technique will be developed that combines depth camera, radar, and lidar sensors’ data. This technique will solve the problem of bad weather conditions by focusing on the sensor that works better in different weather conditions. In order to adapt the vehicle to different road conditions, a benchmark will be developed that creates lots of different and challenging routes with different traffic conditions. The vehicle model will be trained many times in this benchmark. The benchmark will be created on an open-source simulator for autonomous driving research called Carla which we will be using also for testing our algorithms.
Carla has also an autonomous driving leaderboard. The teams train their agent and submit them to the leaderboard. In our project, firstly we will study the submissions on this leaderboard and find out their successful and unsuccessful aspects. After that, we will optimize and use their successful aspects by coding them.
In the first stage, this project aims to work well in the Carla simulator benchmark and rank first in the leaderboard. This leaderboard hosts precious works of researchers trying to bring unique methods for autonomous driving. After being successful in the simulator, the next stage for the project is to apply the developed methods and algorithms to real life. An unmanned ground vehicle will be designed and equipped with a depth camera, lidar and radar. The ultimate purpose of this project is that after ranking first in the leaderboard to see that algorithm is successful, maintaining this success in real life. Furthermore, There are some limitations for this project. One of them is that most of the open-source autonomous driving algorithms do not have evaluation metrics based on the real world. Transition from simulation to real life may be challenging. The other is that scenarios, routes etc. in most of the benchmark tests cannot be changed. Therefore we lack observing algorithms true performance on desired scenarios such as sensor failure. Thus, we will create our own benchmark to overcome these issues.
Otonom araçlar konusunda birçok araştırma yapılmış ve makale yazılmıştır. Bu makaleler arasında en çok elle tutulan ve somut veriler sunanlar Carla liderlik tablosu altında yazılmış olan makalelerdir. Bu liderlik tablosundaki makaleler incelendiğinde kümülatif olarak ilerlenen çalışmanın sonuçları ve otonom araçlar için gelinen son aşama görülebilir.
LAV Literature Review
The third-ranked article in the Carla leaderboard is Learning From All Vehicles[EK-1] by Dotchen et al. In this study, the method is parsed into three steps. The first step, which is called the perception module, extracts the semantic segmentation scores of RGB images and then concatenates this information to the LiDAR point cloud with a sensor fusion approach called PointPainting. Using methods called PointPillars and CenterPoint, 3D object detection is applied on the fused data and a 3D backbone is created. This 3D backbone provides a map-view feature representation by producing detection and semantic mapping. This map-view feature is vehicle-invariant, which means it gives the same features for both the main vehicle and surrounding vehicles.
In the second step a privileged motion planner is created which uses the map-view features of the perception model to produce a series of waypoints describing the future trajectory of the vehicles. This motion planner learns not just from the main vehicle but also from the surrounding vehicles and this is the core point this study suggests that this is what makes them better than other approaches. The fact that the perception module gives vehicle-invariant features makes this approach possible. The motion planner uses standard RNN to predict future waypoints.