Traffic density estimation via a multi-level feature fusion network

Traffic density estimation plays a positive role in improving the traffic efficiency of contemporary cities. Accurate estimation of traffic flow density can provide effective information for traffic dispatching command and effectively alleviate road traffic congestion. However, due to the problems of perspective distortion, scale change, serious occlusion and background interference in traffic video images, it brings great challenges to traffic density estimation. To solve the above problems, this paper constructs a traffic density estimation network based on multi-level fusion network (MFNet). Firstly, the low-level feature map and high-level feature map after Depthwide convolution Block (DCB) are combined to fuse features of different scales, which solves the problems of perspective distortion and scale change in the image; Then, the fused feature map is sent to the channel attention mechanism module to realize the smooth transition between pixels; Finally, by restoring the location information of the vehicle space on the density map and combining with the estimated number of vehicles, the traffic density is calculated quantitatively. In addition, our network also uses multiple superimposed dilated convolutions to obtain high-quality density map. Experimental results show that the Gride Average Mean absolute Error (GAME) metric of the proposed method is reduced to 14.32 on the TRANCOS dataset. Compared with the existing traffic density estimation methods, the estimation accuracy is significantly improved, especially in the case of serious perspective distortion and vehicle height overlap.