Semantic map-based pedestrian trajectory prediction model for intelligent wheelchairs with an improved hybrid LSTM

With the world's population aging rapidly, the demand for wheelchairs as crucial healthcare aids is growing yearly. The safety of wheelchair operation is influenced by the movement trends of surrounding pedestrians. Additionally, due to the unique characteristics of wheelchair drivers, collisions between pedestrians and wheelchairs can pose a high risk of fatality to the driver. Therefore, this paper introduces a novel method for predicting future pedestrian trajectories suitable for intelligent wheelchair driving scenarios. By introducing additional semantic maps, the spatial environmental information in pedestrian trajectory features is effectively enhanced, thereby improving the accuracy of trajectory prediction models. In order to effectively extract feature information from semantic maps and historical trajectories, a feature encoding module based on improved hybrid LSTM is designed. On this basis, a feature fusion and interaction module inspired by attention mechanisms is proposed, enhancing the trajectory prediction model's ability to integrate features from multimodal data. The experimental results indicate that the pedestrian future trajectory prediction method proposed in this paper achieves an average displacement error of 10.17pixels and a final displacement error of 5.95pixels in the image coordinate system, significantly lower than other models. The model proposed in this paper can provide necessary driving assistance for intelligent wheelchairs, thereby improving driving safety.