There have a large number of pedestrian-vehicle accidents on the pedestrian crossing area in China every year, causing huge loss of life and property. In view of different road conditions, it's crucial to establish a more accurate crossing intention recognition model to improve the safety of pedestrians. In this work, a pedestrian crossing area was chosen. Due to construction reasons, two road conditions appeared in the same crossing area at different periods, namely a condition with a zebra crossing and that without a zebra crossing. We compared pedestrian crossing intention parameters under two road conditions in the same crossing area. The results found that there was a great difference in the characterization parameters of pedestrian crossing intention when the site with and without a zebra crossing. Additionally, a more comprehensive crossing intention characteristic parameters set was established. The characteristic parameters were pedestrian speed, the distance between vehicle and crossing area, time to collision (TTC), and safe vehicle deceleration (SVD), pedestrian age, pedestrian gender, group, respectively. The pedestrian intention recognition model for the site with a and without a zebra crossing were established by long short-term memory network integrated with the attention mechanism (AT-LSTM). When the model recognized pedestrian crossing intention 0.6 seconds in advance, the recognition accuracies were 93.05% and 93.89% respectively. The research results are of great significance for improving the safety of autonomous vehicles in the future, and there are also important to improve pedestrian safety.
Read full abstract