针对Tiny YOLOV3目标检测算法在实时检测中对行人等小目标漏检率高的问题, 对该算法的特征提取网络、预测网络、损失函数等进行研究改进。首先, 在特征提取网络中增加2步长的卷积层, 代替原网络中的最大池化层进行下采样; 接着, 使用深度可分离卷积构造反残差块替换传统卷积, 降低模型尺寸和参数量, 增加高维特征提取; 然后, 在原网络两尺度预测的基础上增加一尺度, 形成三尺度预测; 最后, 对损失函数中的边界框位置误差项进行优化。实验结果表明, 改进后的Tiny YOLOV3算法的目标检测准确率比原算法提高了9.8%, 满足实时性要求, 具有一定鲁棒性。本文方法能够更好地提取目标特征, 多尺度预测和边界框位置误差的改进能更准确地对目标进行检测。

The Tiny YOLOV3 target detection algorithm has a high error rate for small targets, such as pedestrians, in real-time detection.Therefore, this study aimed to improve the feature extraction network, prediction network, and loss function of the algorithm.First, a two-step convolution layer was added to the feature extraction network to replace the maximum pooling layer in the original network for downsampling. Second, the traditional convolution was replaced with an anti-residual block constructed by a deep convolutional convolution to reduce the model size as well as number of parameters and increase the high-dimensional feature extraction.Third, based on the original two-scale prediction of the network, a scale was added to form a three-scale prediction. Finally, the boundary box position error in the loss function was optimized.The experimental results demonstrat that the improved Tiny YOLOV3 algorithm achieve a target detection accuracy that IS 9.8% higher than the original algorithm, satisfied the real-time requirement, and demonstratedrobustness.The proposed method can better extract target features, and the multi-scale prediction and improvement of the boundary box position error can detect targets more accurately.