Detection of pedestrians in a crowded scene is a difficult problem since pedestrians usually gather and occlude each other. In addition, due to the complexity of crowded scenes, current deep learning-based approaches for pedestrian detection still require high computational cost. This paper addresses above problems by introducing a deep learning-based approach for fast and accurate pedestrian detection in a crowded scene. To reduce computational cost and increase inference speed, a reduced ShuffleNet network based on ShuffleNet architecture is first adopted as the base network to generate the base convolution layers. ShuffleNet architecture is built on ShuffleNet units and Strided ShuffleNet units, which include pointwise group convolution layers and channel shuffle operations to greatly reduce computation cost while maintaining detection accuracy. To solve the issue of highly overlapped pedestrian in crowded scenes, an improved nonmaximum suppression algorithm is developed based on density score map generated by density prediction sub-network. The improved non-maximum suppression algorithm proposes a dynamic suppression strategy, where the threshold value for suppression rises as pedestrian instances gather and occlude each other and decays when pedestrian instances appear separately. Experimental results on CityPersons dataset and CrowdHuman dataset show the effectiveness of the proposed approach on pedestrian detection in crowded scenes.