Object Tracking Method For Vehicle Mounted Thermal Imaging Pedestrian Detection

Vehicle mounted thermal imaging pedestrian detection is a popular research in computer vision and provides important technical support for ADAS(Advanced driver assistance systems).Thermal imaging is not affected by light,but the resolution,contrast,and signal-to-noise are relatively low.It is difficult to detect objects stably in vehicle mounted thermal imaging pedestrian detection applications.Tracking method helps improve the stability of detection results.Therefore,this paper proposes a classification-assisted KCF(Kernel Correlation Filter)tracking method and applies it to vehicle mounted thermal-infrared pedestrian detection.The main contributions are as follows:(1)When it occurs occlusion or object scale changes,KCF is prone to tracking drift.We use confidence check and classifier check to reduce the tracking drift by calculating the credibility of the KCF prediction bounding box.To improve the calculation efficiency of scale pool strategy,we enable KCF to track multiple scale objects in parallel,and after determine the scale change direction we ignore the calculation in another scale change direction.Experiment result shows that the proposed method improves the tracking precision and success rate by 14.9% and 29.6%,respectively.(2)The detection rate of vehicle mounted thermal imaging pedestrian detection is unsatisfactory.We propose a classification-assisted KCF tracking method.Design Ro Is(Region of Interests)detection list to store Ro Is,and match the Ro Is in the list with the input Ro Is in current frame based on the position and scale of the Ro Is,so as to update the Ro Is in the list quickly.KCF only tracks the Ro Is classified as objects in the list,so it can improve the tracking performance and computing efficiency at the same time.The experiment result shows that the proposed method improves the recall rate by 11.7%.(3)Traffic scene is complex and the influencing factors of thermal imaging are variable.In vehicle mounted thermal imaging pedestrian detection,false positives are prominent and missed detections often result in unstable detection results.We propose a multi-object tracking method.We apply multi-object tracking method into detection system to improve the stability of detecting results and reduce false positives: design a Hungarian correlation algorithm based on Io U(Intersection over Union)and motion direction,and use KCF to complete the correlation between the trajectories and the detected objects.(4)Establish a vehicle mounted thermal imaging multi-object tracking dataset,and verify the effectiveness of the proposed method(3)on this dataset.The experiment result shows that the precision and recall rate are increased by 1.7% and 0.5%,respectively.