Research And Implementation Of Improved Method For Far-infrared Pedestrian Detection On Board

Thermal-infrared pedestrian detection on board is one of the key technologies of advanced driver assistance systems,and it is also a popular research in computer vision.Thermal imaging is not affected by ambient light conditions so it is robust to low visibility and has received extensive attention.Due to dynamics and complex background,varied appearance of pedestrian,and lack of detailed information,thermal-infrared pedestrian detection is still challenging for trading off detection rate and real-time performance.Compared with thermal imaging,many representative methods have been proposed for visible target detection in recent years.The Edge Box method for RoIs extraction has achieved breakthroughs in both recall rate and realtime performance.It uses edge information as a feature and has the potential to be applied to thermal image.The classification stage effectively improves the detection rate and real-time performance by improving visible samples.With the characteristics of the thermal imaging system,the characteristics of the pedestrian target and background on the traffic scene,and the research on visible target detection,this article addresses the requirements of real-time performance and detection rates for thermal-infrared pedestrian detection on board.It proposes a Ro Is extraction method for improving real-time performance and another for improving recall rate,and a detection method based on sample enhancement.The main work of this article is as follows:The EdgeBox method has high recall rate in visible and thermal image,but its real-time performance is still insufficient.So this article proposes a fast Ro Is extraction method named Fast-EdgeBox.Based on the different characteristics of visible common goals and thermalinfrared pedestrians,a cascaded strategy is proposed including pedestrian's size constraints and adaptive local dual threshold segmentation method,to filter the bounding box of background preferentially from the sliding window's traversal phase of the Edge Box.Experiments show that the effectiveness of the improved method,it can effectively reduce the computational cost and the quantity of Ro Is,while maintaining the high recall rate of the original method.When the EdgeBox method is applied to thermal image,the recall rate has decreased.A Ro Is extraction method for improving the recall rate is proposed.By comparing the differences of visible and thermal imaging sparse edge response maps,the reason why the EdgeBox method reduces recall rate in thermal image is analyzed.Based on the Fast-EdgeBox method,a strategy for enhancing vertical edges is proposed to judge the possible pedestrian vertical edges in image,and enhance them.And the method proposes a strategy for improving score evaluation.Based on pedestrian's outline characteristics,it uses a T template to score the filtered bounding box,retaining pedestrian's leg information when rejecting unrelated edges inside the bounding box.A Ro Is reordering strategy based on strongly-constrained size is also proposed to improve the rankings of Ro Is including pedestrian.Experiments show that the improved method can effectively improve the recall rate,while the real-time performance is slightly reduced.Aiming at the requirements of balance the real-time performance and the detection rate on the classification stage,a detection method based on sample enhancement is proposed.The article analyzes the existing problems of the thermal imaging samples,and proposes strategies for improving sample in terms of quantity and data distribution,then generate positive and negative enhanced samples,and uses the enhanced samples to train the Classifier with ACFT+THOG and AdaBoost.Experiments show that the improved method can effectively reduce the missed detection rate compared to the same type of method using the original sample.