The Density-Aware Estimation Network for Vehicle Counting in Traffic Surveillance System

In a surveillance system, the performance of vehicle counting is typically sensitive to the different patterns of traffic density. To address this issue, we proposed the estimation network utilizing intermediate layers for mapping between the appearance of the input image and its density map, where each layer corresponds to different traffic density. The backward connection is introduced by extracting feature maps from a deeper layer to combine with a shallower layer. Taking advantage of the high-level feature from deep layers, more informative features are extracted in the shallow layers, where the performance in an earlier stage can be improved. The experimental results show that the proposed method provides more stable performance in various traffic density than traditional approaches, where the difference between the maximum and the minimum errors in various density is about 5, evaluated by mean absolute error.