Forward Collision Warning (FCW) system is crucial not only for the implementation of autonomous vehicles but also for reducing traffic accidents. According to the statistics of the United States and Germany, the number of rear-end collisions during th...
Forward Collision Warning (FCW) system is crucial not only for the implementation of autonomous vehicles but also for reducing traffic accidents. According to the statistics of the United States and Germany, the number of rear-end collisions during the entire accident was close to 29.5% in the United States and 28% in Germany. Also, according to the Korea Transport Institute (KOTI) statistics, 98.6% of road traffic fatalities involve human factors, and according to the World Road Association (PIARC) data, it is investigated that 93% of traffic accidents are related to human factors. According to a study by Daimler-Benz, if the driver responds in advance of 0.5 seconds, it can prevent 60% accident, and if it responded in advance of 1 second, it would prevent 90% of accidents.
However, an additional option for advanced driver assistance systems with equipment such as camera, rider sensor and radar sensor should be added at a high additional cost to use the forward collision warning system on most of the finished vehicles. In this paper, I propose FCW system based a single camera using deep running and OBD-II without the use of radar sensors or stereo cameras generally used for distance measurement. The proposed method increases object detection rate using YOLO (You Only Look Once) deep learning network and is consist of steps of object detection, distance measurement, TTC (Time to Contact) calculation and forward collision warning. In the object detection step, an object with potential collision during driving is detected through the YOLO network and the vehicle in front of the driving vehicle is tracked. In the distance measurement step, the distance to object with possible collision is calculated through the proportional expression between the camera and the real object. In the TTC calculation step, TTC is calculated through the difference between video frames. At this time, information of OBD-II is used as an auxiliary index. In the forward collision warning step, a collision warning is sounded when the TTC calculated in the previous step falls below the threshold value.
Scale Invariant Feature Transform (SIFT) and Histogram of Oriented Gradient (HOG), which are based on existing vision processing, generate errors of about 30% because of calculating the vehicle width by detecting object corners. On the other hand, when the object detection is performed by YOLO network, the error of the detected object width is reduced to about 10% by detecting the bus, truck, and passenger car separately. Thus, it is possible to estimate the TTC based on the distance calculated using the characteristic of the camera. The collision test result based on the actual calculated TTC shows that the forward collision warning system operates only when there is a possibility of actual collision by tracking the object in front of driving direction of the own vehicle. Also, in accelerating and constant-speed collision experiment, the forward collision warning is sounded 2.5 seconds before the actual collision, helping the driver to brake before the collision.