Optimization Method For Vehicle-mounted Thermal Imaging Pedestrian Detection

As the number of cars in our country continues to grow,the problem of driving safety becomes more and more serious.The vehicle-mounted thermal imaging pedestrian detection system detects the pedestrians in the road in real time,warning the driver to avoid collision,and further reduce the accidents effectively.However,due to the lack of color information,target details and texture,and pedestrians in the real traffic scene usually encounter many interferences such as occlusion,uneven heat distribution and large scale difference,it is difficult to detect pedestrian targets in thermal image accurately.Based on the characteristics of pedestrian in thermal image,this paper studies the existing pedestrian detection algorithm and proposes optimization methods to improve the detection performance.The main contributions are summarized as follows:1)The head pattern of pedestrian in thermal image is stable.Therefore,we propose a region of interest(Ro Is)extraction algorithm based on head key points for vehicle-mounted thermal imaging pedestrian detection.Using “frame head region” and “foreground region” to restrain the head key points extraction so as to improve its accuracy and efficiency.Besides,we design an efficient Ro Is model to calculate the head key points to generate Ro Is quickly,and use foreground information to adjust Ro Is to improve the location accuracy.The experimental results show that the proposed method is obviously better than the segmentation-based method on the coverage rate and recall rate under the premise of ensuring efficiency.2)Unbalanced sample distribution and low efficiency are common problems in vehicle-mounted thermal imaging pedestrian detection.Therefore,we propose a classifier optimization method based on HOG-XGBoost for vehicle-mounted thermal imaging pedestrian detection.Applying Focal-Loss into XGBoost training process to solve the imbalance of sample distribution and improve the accuracy of the classifier.And we design a strategy based on the embedded feature selection method to construct the HOG feature extraction template.It can effectively improve the speed of feature extraction,and furtherimprove the efficiency of the classifier.Experimental results show that the proposed pedestrian detection system can achieve high detection rate in various traffic scenes.3)The heads of pedestrians can represent pedestrians stably in thermal image.Therefore,we propose a thermal imaging pedestrian detection method with head enhancement module based on deep learning.We design the head key point regression enhancement module and head region segmentation enhancement module to make the neural network fully learn the semantic relationship between the head and the whole pedestrian,so as to improve the pedestrian detection performance.Besides,we design an automatic head annotation method to provide head ground truth for training the neural network effectively on the dataset without accurate manual head annotations.Experimental results show that the proposed method can effectively improve the detection performance without increasing any inference time.