Vehicle Crowd Density Estimation Enhanced by Video Flow Maps

This paper proposes an approach to vehicle crowd analysis in video sequences that utilizes flow map extraction based on density maps. We adapted the CANnet2s deep neural network, originally designed for human crowd analysis, to the vehicle domain. In particular we trained CANnet2s from scratch using Stochastic Gradient Descent, where VGG16 backbone trained on ImageNet is used as the pretrained model. Our primary motivation is enhancing vehicle crowd analysis for autonomous driving, hence evaluation of the network is performed on Waymo dataset. We annotated Waymo data with seven attributes, such as occlusion, pose change, bright, dark, blurry, multi-scale, and low density. Performance is reported by MAE, RMSE, and PSNR metrics. Comparative analysis with CANnet, the still-image-based counterpart of CANnet2s, demonstrate that the proposed approach outperforms the still image based learning at all attributes and achieves significant improvement in blurry scenes.