基于Jetson TX1的目标检测系统

针对目标检测算法在实际应用中速度仍需提高的问题,提出一种改进的YOLOv3算法.通过将多层次特征金字塔网络嵌入到YOLOv3算法中的DarkNet-53网络,解决了由实际检测对象尺度差异引起的误差问题,既增加了输出特征的尺度,又加深了网络层数;并针对损失函数中的激活函数零均值、梯度消失等问题,提出了一种新的激活函数.实验结果表明,改进的YOLOv3算法可以更好地对多尺度目标进行检测,提高了检测速度和准确性,并在硬件上实现时具有较为高效的性能表现.

Object detection system based on Jetson TX1

Aiming at the problem that the object detection algorithm still needs to be accelerated in practical application, an improved YOLOv3 algorithm was proposed. Through embedding the multi-level feature pyramid network into the DarkNet-53 network of YOLOv3 algorithm, the error problem caused by the scale difference of actual detection objects was solved. The as-proposed algorithm not only increased the scale of output features, but also made the number of network layers deepened. Concerning the problems of zero mean of activation function and vanishing gradient in loss function, a novel activation function was proposed. The results show that the improved YOLOv3 algorithm can detect the multi-scale objects much better, improve the detection speed and accuracy, and exhibit a more efficient performance for hardware implementation.