DeGOA: an optimized deep LSTM model for vehicle trajectory prediction

Engineering notes

Key topics: autonomous driving, trajectory prediction, prediction, planning. See the paper for implementation details and experimental results.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

The ability to predict the future position of the vehicle is decisive for various applications like autonomous driving, intelligent transportation systems (ITS), and urban mobility planning, as it allows the user to generate the future vehicle motion given the state of surrounding vehicles, their environment, and more. The paper presents a deep learning model that leverages the SmoothGolf algorithm for optimizing vehicle trajectory prediction by reducing error rates and improving robustness. The first step consists of extracting some properties, like entropy, tortuosity, elongation, localization, and elongation angle from the input trajectory data, as each describes some aspect of the vehicle movement dynamics. The second step involves feature selection using techniques such as the Tversky index, Pearson correlation, and cosine similarity, each enhancing data relevance through different criteria. The essence of the third stage is the use of a deep long short-term memory (LSTM) network for prediction, which has the capability of learning long-term dependencies that play a significant role in predicting vehicle trajectories. Here, DeGOA is proposed, which combines double exponential smoothing (DES) with golf optimization algorithm (GOA) for optimally adjusting the parameters of the deep LSTM. The DES with mean value suppression helps to smooth the trend in the temporal information, thereby enhancing prediction accuracy in the presence of noise, and the swarm intelligence-based GOA assists in performing parameter searches effectively. Experimental analysis shows that the SmoothGolf-based deep LSTM is highly effective in generating accurate trajectory predictions. The results of the MSE of 0.001, RME of 0.044, and MAE of 0.008—validate the proposed method’s robustness and accuracy. The results exhibit their potential for generating reliable and optimal vehicle trajectory predictions.

Links and sources

Comments