关键词: 深度学习, 卷积神经网络, 图像分类, 行人检测

Abstract: In order to make the convolutional neural network in pedestrian detection to choose better model and acquire more accurate bounding-boxes, an improved pedestrian detection method based on convolutional neural network is proposed. The improvements include two aspects: how to find a better iterative number of the training samples and how to merge multiple responses of an object. On the solution of the first improvement, multiple candidate CNN classifiers are learned from different training samples in different training iterations.. And a new strategy is proposed to select the model with better generalization ability. The strategy considers both the accuracy on the validation set and the stability of the accuracies during the iterative training procedure. On the improvement of combining multiple responses, an enhanced refined bounding-box combination method is proposed which is different from the non-maximum suppression method. The method takes the coarse bounding-boxes obtained by the CNN detection procedure as input, then gets the one-to-one refined bounding-box obtained by a CNN refinement procedure. And finally merges multiple refined bounding-boxes considering the bounding-box correction probability. Exactly, the final output bounding-box is the weighted average of multiple relevant refined bounding boxes with respect to their correction probabilities. To investigate the proposed two techniques comprehensive experiments are conducted on well-recognized pedestrian detection benchmark dataset (ETH). The experimental results show that the two proposed techniques are both effective under the same evaluation protocol. Comparing with the benchmark method of Fast R-CNN, the average miss rate of the proposed method is greatly reduced with a margin of 5.06 percentage points.

Key words: deep learning, convolutional neural network, image classification, pedestrian detection