Vehicle Crowd Analysis via Transfer Learning
Yusuf K. Hanoglu, Bilge Günsel, Meltem Gulbas · 2023
We propose a deep learning based approach to vehicle density estimation that adopts CSRNet, originally designed for person crowd analysis, to vehicle crowd analysis. The objective is to exploit the transfer learning to accurately estimate the vehicle density with an increased learning speed. Specifically, the CSRNet architecture pre-trained on the person domain is fine tuned on the vehicle domain by feature tranformation. This is achieved by end-to-end retraining the network to output the spatial distribution of vehicles in congested scenes. The approach is evaluated on Waymo and TRANCOS data sets while ShanghaiTech data set is used for pretraining. Performance reported by the metrics of MAE and RMSE, and PSNR on different test cases, demonstrate the transfer learning significantly improves vehicle density estimation accuracy, compared to the learning from stretch. In particular, the learning accuracy achieved on Waymo, with a small size training data, is validating the potential of the approach in enhancing vehicle crowd analysis for autonomous driving task.