Unbalanced taxi demand and supply always cause not only passengers to wait for a long time, but also the taxi company to rise their operation cost. To address this problem, many researchers are focusing on predicting taxi demand of areas in the city as precisely as possible by mining various types of data (e.g., weather, events, etc.). However, prediction models can only get the predicted result in a rough area that cannot guide taxi drivers accurately. On the other hand, predicted result may not be always precise due to the changeful situations in the real-world.
In this thesis, we propose a vision based system to improve this problem. In our approach, we detect the taxi demand hotspots through processing the streaming video data from taxi's drive recorder, and analyze the status of pedestrians via computer vision technologies to estimate whether they are waiting for taxis. All of our system is deployed on Nvidia Xavier mobile computing platform which install on taxis. So this information of where may exist potential pedestrians will be shared among taxis in real-time to help vacant taxis pick up passengers.
Specifically, our proposed method consists of four steps as follows. We firstly detect pedestrians appearing in the streaming video data. Subsequently, detected pedestrians are tracked through Multi-object tracking technology to get their temporal patches. Then we feed the patches to an action recognizer to distinguish whether he/she is walking or standing. Finally, we only analyze the status of standing pedestrians to get a probability score of how likely they are waiting for taxis. Besides, we also implement a zebra crossing detector to shut out standing pedestrians who are waiting for traffic light rather than taxis.
To train and verify our system, we created a dataset which contains 19 participants acting different status (e.g., waving hand to taxis). In evaluation, we followed k-fold cross validation (k=5) manner, and got result of 87.16% Accuracy and 0.802 F1 score of detecting waiting-taxi pedestrians with 53.53 ms inference time. Because the visual information can most intuitively reflect the real situation of streets, our approach provides a more direct solution to help taxi drivers find potential passengers efficiently in real-time and more robustly to the changeful situation.