Pedestrian detection method in vehicle video based on AMSSD

Pedestrian detection in vehicle video has high requirements for speed and accuracy of detection. To improve the detection speed, the underlying network of the SSD framework, the VGG-16 network, is replaced with MobileNet. In order to improve the detection accuracy of small targets, a deconvolution layer is added to this framework. With this layer, the feature map of low resolution and high semantic information with the feature map of high resolution and little semantic information is fused to increase the ability to extract the shallow feature. The VOC dataset and COCO dataset are used as the training set, and the Cityscapes dataset is used as the test set to verify the effect of the constructed framework AMSSD. The experimental results show that the proposed method can improve the detection speed and accuracy.