基于Tiny-yolov3的行人目标检测研究

Research on Pedestrian Target Detection Based on Tiny-yolov3

为了改善传统行人检测算法鲁棒性差、检测精度不高、实时性差、训练环境对硬件依赖性强的缺点,基于Darknet框架,使用Tiny-yolov3目标检测模型,在原有模型基础上,通过强化行人特征、改进算法网络结构,并不断调整学习率、动量、权重衰减系数等超参数,将模型放在混合数据集上进行实验。实验结果表明,基于Tiny-yolov3的改进行人目标检测模型准确率、召回率、平均交并比和FPS值较高,分别达81.13%、76%、83.76%和62帧/s。基于Tiny-yolov3的改进行人目标检测模型不仅能对实际场景进行更实时的检测,而且有效降低了模型对硬件的依赖,同时也能提高行人目标检测准确度,提升识别效率。

In order to improve the shortcomings of traditional pedestrian detection algorithms,such as poor robustness,low detection accuracy,poor real-time performance,and strong dependence on hardware in the training environment,based on the Darknet framework,the Tiny-yolov3 target detection model is used.On the basis of the original model,the model is placed in the mixed data set by strengthening the pedestrian characteristics,improving the algorithm network structure,and constantly adjusting the learning rate,momentum,weight attenuation coefficient and other hyperparameters conduct experiment.The experimental results show that the improved human target detection model based on Tiny-yolov3 has higher accuracy,recall,average cross-to-bin ratio and FPS value,reaching 81.13%,76%,83.76%and 62 frames/s,respectively.The modified human target detection model based on Tiny-yolov3 can not only perform more real-time detection of actual scenes,but also effectively reduce the model’s dependence on hardware.At the same time,it can also improve the accuracy of pedestrian target detection and improve recognition efficiency.