Abstract: Current methods of extracting regions of interest are difficult to balance the high recall rate and the number of ＲoIs． To reduce the amount of computation and ＲoI，and increase recall rate，a ＲoIs extraction method suitable for robust vehicle thermal imaging pedestrian detection was proposed． First，the vertical edges of possible pedestrians in the image were judged according to the directional differences of pedestrian edge features，and their amplitude was enhanced． Then，the boundary boxes generated by sliding windows were filtered by cascade human dimension constraint and adaptive local double threshold segmentation method，and a large number of non-human boundary boxes were removed． Next，according to the outline characteristics of pedestrian，the filtered bounding boxes’score were evaluated by using a T-shaped template，and the unrelated edges inside the bounding boxes were removed while preserving possible pedestrians’leg information． Finally，the ＲoIs were reordered with the strong size constraint of pedestrians，so that the recall rate could be improved when a fixed amount of ＲoI was extracted． A contrast experiment was conducted on a thermal imaging pedestrian detection dataset called SCUT DataSet． The results show that when 400 ＲoIs are extracted from each image，the recall rate of our method reaches 92%; compared with the EdgeBox method，the recall rate increases by 21%，and the calculation time decreases by 10%．

Key words:  vehicle thermal imaging, pedestrian detection, regions of interest extraction, local double threshold segmentation, pedestrian safety, edge detection