Investiagation On Perception Technology Of FSAC Racecar Based On Monocular Vision

As one of the key research targets of most advanced science,driverless technology has influenced greatly on economic growth,road safety and national defense security.Environment perception is one of the key driverless technologies which determines whether the autonomous vehicle can obtain the comprehensive road information and plan a safe and reasonable driving route.Therefore,it’s of great significance to study the environment perception technology.In the present study,focused on the Formula Student Autonomous China(FSAC),we aimed to solve the current challenges about the dysfunction of sensors caused by the light alternation,which will result in collision and deviation of the vehicle.We proposed a technical solution that based on the monocular vision detection of barrels and colors,allowing the driverless vehicles to recognize barrels in distinct colors and arrays,under different lighting conditions.We first investigated into the rules and environmental parameters of the formula racing.Based on the results,the cameras were able to be selected and installed.Meanwhile,we conducted visualization and subpixel-level precision of corners in camera calibration which reduced the reprojection error to 0.0417,so as to ensure the reliable results of the camera output.Then,we studied the detection algorithm and proposed to use YOLOv5 algorithm to detect the buckets as single category.On the bucket sample data set constructed in this paper,the training results of 125 fps frame rate,98.31% precision rate and 100% recall rate were obtained,which provided a candidate region of bucket with high confidence for the color recognition algorithm.On this basis,in order to eliminate the redundant information in the candidate region of bucket,we studied the image processing algorithm and proposed the object segmentation method which separates the channels of hue,saturation and value.Meanwhile,we proposed to use different times of closing operation for buckets with different colors and use median filter for noise reduction.At last,we presented an algorithm that based on HSV color-space discernment,which transformed the discrimination problem of colors into the comparison problem of contour area sizes,allowing precise recognition of the buckets in the confined area in which only valid information is kept.The results of this research have innovatively combined the novel deep-learning with traditional color recognition,avoiding the problem of sample imbalance while preserving the interpretability.In the off-line simulation,the accuracy of color recognition under strong,normal and dim light is 98.87%,99.06% and 98.56%,respectively,on the basis of the frame rate of 27 fps,which indicates that this paper effectively solves the safety problems caused by light changes in the FSAC racing environment awareness module.