基于单目视觉的小目标行人检测与测距研究

Study on Small Target Pedestrian Detection and Ranging Based on Monocular Vision

自动驾驶场景下,为了提高远距离行人的检测精度和测距精度,结合深度学习的目标检测,提出了一种行人测距算法。首先,提出了冗余切图法,结合YOLOV3模型对小目标行人进行检测,再通过改进的边界框筛选算法对所有子图的候选框进行多次筛选,最终得到行人检测框。然后,对传统的相似三角形测距算法进行分析,提出了一种包含pitch和yaw的改进相似三角形测距算法。最后,根据行人检测结果实时测量行人距离当前车辆的横向距离和纵向距离。实验结果表明,在BDD 100K验证集上,所提出的冗余切图法检测模型比原YOLOV3模型的mAP提高了6%,比小目标行人的mAP提高了3%,具有更好的检测鲁棒性;在车载摄像头采集的测距测试集上,冗余切图法和改进的测距算法的结合使测距精度在对比实验结果上改善了6.542%,不仅实现了远距离测距,而且具有更高的测距准确性。

In order to improve the detection and ranging accuracy of long-distance pedestrians under automatic driving scenario,a pedestrian ranging algorithm is proposed based on the target detection of deep learning.Firstly,a redundant graph cutting method is proposed to detect the small target pedestrian by combining with the YOLOV3 model.And then the candidate bounds of all subgraphs are screened many times by the improved bounding box screening algorithm,and finally the pedestrian detection box is obtained.By analyzing the traditional similar triangle ranging algorithm,an improved similar triangle ranging algorithm including pitch and yaw is proposed.Finally,the transverse and longitudinal distances between pedestrian and the current vehicle are mea-sured in real time according to the pedestrian detection results.The experimental results show that,on the validation set BDD 100 K,the mAP of the proposed redundant graph cutting detection model is 6%higher than that of the original YOLOV3 model,and the mAP of small target pedestrian is improved by 3%,and has a better robustness.On the ranging test set collected by on-board camera,the combination of the redundant graph cutting method and the improved ranging algorithm improves the range measurement accuracy by 6.542%compared with the experimental results,which not only realizes long-distance range measurement,but also has higher range measurement accuracy.