In smart cities, traffic management is becoming a significant challenge owing to the rapid growth of population. In this regard, pedestrian detection and their safety are of the utmost importance for the city authorities. In this paper, a fully convolutional deep architecture is presented to detect pedestrians by automatically selecting the desired region proposals as well as learning the requisite feature representation with no need for any manual hand-crafted feature design. The architecture facilitates end-to-end training and thereby improves the overall performance, eliminating the bottleneck caused by the multistage pipeline structure of conventional feature engineering. A state-of-the-art deep framework, for general object detection, is suitably tailored for the task of pedestrian detection. A densely connected convolutional network is employed to learn the desired features. A two-stage approach is proposed to separate the human-look-alike hard negative backgrounds from the true pedestrians. Besides, feature maps from multiple intermediate layers are taken into consideration to facilitate small-scale detection. The proposed method alongside a few competent schemes is compared on the benchmark Caltech and INRIA datasets, where the log average miss rate is set as the performance metric. The obtained results demonstrate the potential of our approach in addressing the real-world challenges. © 2018 John Wiley & Sons, Ltd.