Pedestrian Tracking and Predicting Methods in Traffic Roads

In order to analyze the movement intention of pedestrians in traffic roads, this paper proposes a pedestrian tracking and predicting method incorporating fused attention. The method uses CenterTrack to track pedestrians, adds fused attention to its DLA34 to enhance the extraction of key features of the tracked objects and improve the efficiency of information transfer in the network, representing the extracted attributes of the tracked objects as points on a heat map and correlates them with a greedy matching strategy to track the objects accurately, then uses the LSTM to loop through the historical locations of the tracked objects to predict their future locations, and finally visualizes the tracked and predicted locations. The experiments show that the improved CenterTrack outperforms other tracking algorithms in many aspects, with MOTA reaching 91.9 and MOTP reaching 80.3; the predicted coordinate errors of LSTM for pedestrians are within a small range, with the highest MAE of 6.13px and the highest RMSE of 8.26px in 1280 × 720 video; combined with the visualization results. The experimental data and visualization results demonstrate that this method can effectively accomplish tracking and predicting the location of pedestrians in road conditions.