Research On Detection Of Infrared Thermal Imaging Target Under Poor Driving Vision

Safe driving in traffic has been a hot topic of research in the field of transportation for a long time.This paper studies the infrared thermal imaging target detection under poor driving vision,aiming to deal with the safety problem under bad conditions of darkness,smoke,bad weather and glare.Conventional target detection recognition lacks real-time capability,presents fewer black and white features in infrared images,recognizes poorly in the multi-scale target.Deep learning target detection algorithms can meet the balance of accuracy and speed with powerful computational learning features.This paper investigates target detection based on the YOLOv4 algorithm in a convolutional neural network for infrared thermal imaging,which involves the following procedures:(1)The relevant literature is consulted to analyze the importance of choosing infrared thermal imaging for target detection,to study the computational principles of image localization and classification,and to find an effective deep learning target detection method.(2)A test target detection test program is established in the context of poor driving vision to clean data and converse to the standard data format for the defects of the FLIR infrared thermal data set.Constructed an infrared thermal imaging target detection platform,a detection model is trained using a migration learning approach to fine-tune the data.In the measured results,the self-training model has a better improvement in accuracy than the original YOLOv4 model,and there is no obvious improvement in speed.The AP of the model to detect pedestrians and cars were 86.83% and 90.72%,respectively;the overall was Avg IOU 65.17% and m AP 88.8%.The FPS of the self-training model for real-time video detection is about 35.5,and the video scene is the city road when high population flow at night.From the measured detection effect,the target frame localization is of high accuracy,and the prediction result is satisfactory.Meanwhile,the real-time result playback is smooth,which can meet the safety requirements of assisted driving.(3)The algorithm is improved by anchorbox clustering center according to k-means and k-means++ methods.Comparing the convergence of Avg IOU after 80 iterations of computation in two sets of clustering experiments with k=5 and k=9,it is verified that kmeans++ is more suitable for the study of algorithm improvement.The model trained by the improved algorithm has the performance statistics of Avg IOU 69.88% and m AP 90%,which are 4.71% and 1.2% higher than before the improvement,respectively.Although the accuracy is improved,the speed on the final real-time video detection remains the same.In the present study,the infrared thermal imaging target detection has excellent results in both speed and accuracy,which can effectively improve the safety of driving under poor vision.