Safe training of traffic assistants for detection of dangerous accidents

Abstract

As the automotive industry continues to develop, the car's function is no longer limited to a simple means of transportation. Instead, it is more of a technological product that combines safe mobility, entertainment and safe driving technology. Furthermore, ensuring the safety of passengers is a critical element in the development of cars. Thus, safety-based autonomous driving and assisted driving systems are essential in developing cars and future development strategies. However, the limitations of the current radar systems used in automobiles are widespread. A single radar detection makes it difficult to carry out accurate object detection and identification and can only provide vague conceptual feedback. The radar feedback needs to be more accurate, especially when the distance is too close or too far. Any object that reflects radar waves is being used as a hazard warning. Due to the continuous development of information technology and artificial intelligence technology, integrating artificial intelligence into traditional industries to achieve automation and intellectual development is the main direction of current technological development and industry progress. For example, the application of AI technology to the automotive industry enables comprehensive and immediate environmental awareness, comprehensive and accurate planning and decision-making, and precise and efficient vehicle control to ensure the safety of passengers. In this thesis we propose the use of the YOLO algorithm in Virtual Worlds to safely train the car's recognition to detect dangerous traffic accident situations in different environments without damage to property and danger to human well-being through real-time video detection by obtaining more accurate information about obstacles or hazards. The YOLO series of Artificial Intelligent (AI) detection algorithms are used to detect objects through video or pictures. Unlike radar detection, YOLO can accurately analyse obstacles. Assisted driving and autonomous driving will be an essential part of the future of transportation, but training them for object detection and recognition of dangerous traffic situations, which is a key aspect of its operation, is difficult because of the damage to property and human well-being. Therefore performing this training in virtual world is essential. First, we designed and built a virtual 3D city platform using the Unity 3D engine, recreated as much realistic road information as possible in the 3D city. Then we used the YOLOV5 algorithm for detection of objects to obtain accurate virtual identification information successfully. After training, YOLOV5 can detect all vehicles and obstacles on the virtual road. From there, it can alert the driver of dangerous traffic situations accordingly instead of alerting for all objects to avoid unnecessary danger warnings. On the other hand, environmental perception is essential to safe driving. Nevertheless, current research has seen various technologies applied to environmental perception, such as Microsoft's AIRSIM autonomous driving simulator, LIDAR technology and millimetre wave radar technology, which are currently heavily used. However, technology is constantly evolving, and LIDAR and millimetre wave radar are now at the forefront of environmental awareness. Accurate one-stage algorithms and databases are an important direction for the future. This is because such algorithms not only indicate the presence of an object in front of them but also identify exactly what type of object the output is(people, pets, ground obstacles, etc.). We have built on the Yolo algorithm and applied it to assisted driving with a focus on safely training the AI to detect the driver's blind spot, by analysing the environment out of the driver's view and giving timely feedback. This thesis explores in depth the application of how to safely train YOLO for assisted driving, building a 3D virtual city and testing it in different stages in a virtual environment. The usefulness of the YOLO algorithm for driving car safety is verified. Through the continuous training of the YOLO algorithm, an extensive database can give the driver more results in terms of environmental perception. As a result, the occurrence of traffic accidents due to insufficient environmental perception for training YOLO can be increased by constructing virtual accidents without damage to property and human well-being.