基于改进YOLOv3的街道行人检测与跟踪方法

针对室外街道的行人检测与跟踪,提出一种改进YOLOv3与简单在线实时跟踪(simple online and real-time tracking,SORT)算法相结合的检测及跟踪方法.首先,引入距离和比例交并比(distance and proportional-IOU,DPIOU)损失,将原有的损失函数中的均方误差(mean square error,MSE)部分进行变化,从而得到更精确的检测框;其次,将网络结构中的RestNet进行优化,改变下采样区域,增加池化层,进而减少特征信息的丢失;最后将检测结果输入SORT算法进行建模和匹配.实验结果表明,在室外街道的场景下,改进的算法与YOLOv3相比较,损失值收敛更快,平均准确率高出4.85％,跟踪准确率上升3.4％,同时,模型的速度有所提高,最快可达14.39 FPS.

Pedestrian Detection and Tracking Method on Street Based on Improved YOLOv3

Aiming at pedestrian detection and tracking in outdoor street, an detection and tracking method based on improved YOLOv3 and SORT(Simple Online And Realtime Tracking) is proposed. Firstly, DPIOU (Distance and Proportional-IOU) loss is introduced to change the MSE (Mean Square Error) part of the original loss function to get a more accurate detection frame; secondly, the RestNet in the network structure is optimized to change the down sampling area and add the pooling layer to reduce the loss of feature information; finally, the detection results are input into SORT algorithm for modeling and matching. The results of experiment show that, compared with YOLOv3, the improved algorithm has faster convergence of loss value, AP value is 4.85 higher, and tracking accuracy rate is increased by 3.4% in the outdoor street scene. Meanwhile, the speed of the model is increased and the maximum speed is 14.39 FPS.