一种轻量级的卷积神经网络的行人检测方法

A PEDESTRIAN DETECTION METHOD BASED ON LIGHTWEIGHT CONVOLUTIONAL NEURAL NETWORK

为了解决传统前端行人检测算法准确率低以及鲁棒性差的问题,提出将前端嵌入式设备与人工智能芯片搭载的轻量级神经网络相结合的方法,以实现在前端嵌入式设备中完成更加准确、稳定的行人检测。针对前端嵌入式设备性能不足的问题,提出一种轻量级卷积神经网络模型,通过对网络框架的重新设计以及使用聚类分析重新定义候选框尺寸的方法,大大减少网络权重以及计算量。实验结果表明:该方法相较于传统行人检测方法有着更高的检测精度,并在嵌入式行人检测加速方面具有显著效果,可达到每幅图片62 ms的检测速度。

In order to solve the problem of low accuracy and poor robustness of the traditional front-end pedestrian detection algorithm,this paper proposes a method combining the front-end embedded device and the lightweight neural network equipped with the artificial intelligence chip to achieve more accurate and stable pedestrian detection in front-end embedded devices.Aiming at the insufficient performance of the front-end embedded devices,a lightweight convolutional neural network model was proposed,and it reduced network weights and computations by redesigning the network framework and redefining candidate frame sizes using cluster analysis.The experimental results show that the method has higher detection accuracy than the traditional pedestrian detection method,and has significant effect on the embedded pedestrian detection acceleration,which can achieve the detection speed of 62ms per picture.