Design Of Thermal Infrared Warning System For Vehicle Collision In Low Visibility

When the vehicle is driving in low visibility environments such as rain,snow,fog or night,the driver ’s line of sight is seriously disturbed,which affects the judgment of the road conditions and vehicle conditions ahead,and often leads to traffic accidents.At present,machine vision detection and recognition technology based on visible light has poor adaptability to the environment.Based on these situations,a thermal infrared vehicle forward collision warning system is studied and designed.The main research works are as follows :(1)The demand for machine vision in low visibility scenes is analyzed,and a vehicle forward collision warning system using thermal infrared detection technology is designed,including target recognition,target ranging and early warning judgment system.In the target recognition system,the target category,pixel size and pixel position in the vehicle scene are obtained in real time and accurately by the detection algorithm.Then,the precise ranging of different targets is carried out,and the data needed for early warning and judgment is obtained.In the early warning,whether it is a dangerous target is judged by combining the target category,relative position and the safety distance and warning time set at the current speed,and the dangerous target is prompted.(2)The thermal infrared vehicle scene data set in low visibility environment is established,and the YOLOV5 target detection algorithm is used to realize the recognition of pedestrians and vehicles in low visibility environment.Aiming at the problem of low accuracy of YOLOV5 target detection algorithm,this thesis improves the anchor point of YOLOV5 algorithm based on the characteristics of data set,and adds a small target detection layer.Compared with the original YOLOV5 target detection algorithm,the improved algorithm mean-Average-Precision increased from 82.1% to 87.9%.Aiming at the problem that the YOLOV5 target detection algorithm can not meet the real-time performance in practical applications,the improved target detection algorithm is deployed to NVIDIA AGX Xavier,and the inference is accelerated by TensorRT.Experiments show that the frame rate of model inference after acceleration is increased from 19 fps to 30 fps.(3)A ranging algorithm based on target detection box is proposed.Aiming at the problem that the traditional monocular ranging algorithm has large ranging error for different targets,the five types of targets are divided into three categories.The distance calibration of different types of targets is carried out respectively.The distance fitting is carried out according to the calibrated data,and the ranging formulas of different types of targets are obtained.Compared with the traditional monocular ranging algorithm,the average error of pedestrians within 50 m is reduced from 4.65 m to 0.46 m.The average error of small cars is reduced from 3.59 m to 0.45 m.The average error of large vehicles is reduced from 0.95 m to 0.43 m.The actual scene operation test of the developed vehicle anti-collision instrument shows that the thermal infrared imaging technology integrating multi-source data can achieve highprecision,low-latency reliability detection and meet the vehicle safety warning requirements in low-visibility scenarios.