New path prediction model for autonomous vehicles in construction zones

The problem of motion prediction in dynamic traffic conditions

Path prediction autonomous vehicles in construction zones is one of the most challenging tasks in the field of mobility. Changes in road layout, temporary crossings, and the presence of workers on the road create a highly uncertain environment. Research shows that this topic is still under-researched - there is a limited amount of specialized data and a lack of systematic evaluation of solutions. In real-world conditions, such zones are often unpredictable, which makes it difficult for autonomous vehicles to make safe decisions.

WZDrive-Data is the first multimodal dataset collected in a human-dependent manner in construction zones. It contains 1,183 driving segments with the participation of 33 participants who drove a vehicle in a simulator with animations of working robots and user-controlled virtual reality. During the experiment, not only the driving path was recorded, but also visual and control signals from the participants, which made it possible to analyze both human behavior and interaction with the environment.

New WZDrive model and its comparative results

Based on the WZDrive-Data dataset, a new driving path prediction model was developed - WZDrive. The model uses a multimodal approach, combining data from human signals (e.g., gaze direction, steering wheel movement) with information about the scene and driving path. Its main goal was to improve the accuracy of prediction in high-risk and variable conditions. In comparative tests, the model achieved an average minADE (K=6) value of 1.09 m and a minFDE (K=6) value of 2.12 m - which means that the predicted paths were on average less than a meter away from the actual path in the case of the most likely trajectories.

WZDrive achieved better results than the best state-of-the-art model so far: reduction of minADE (K=6) by 7.18% and minFDE (K=6) by 13.18%. These numbers show that the model not only improves prediction accuracy, but also significantly reduces the risk of incorrect decisions in difficult conditions. The results are particularly important for safety - even a small difference in predicting the path can affect avoiding a collision with a worker or another vehicle.

Significance and limitations of research

WZDrive research shows that integrating human data into the model learning process can significantly improve their effectiveness in difficult conditions. First of all, this approach allows to take into account not only physical changes in the environment, but also human behavior - which is crucial in construction zones. This may lead to more flexible and safer autonomous driving systems, especially in cities where construction zones are common.

At the same time, it should be remembered that the results were obtained in a simulated environment. The lack of data from real-world road conditions means that the accuracy of the model in the physical world may vary. In addition, the dataset includes only 33 participants and a limited number of scenarios, which may affect its representativeness. Further research requires expanding the dataset and testing it in real-world conditions to assess the scalability and durability of the model.

The future of autonomous driving in complex environments

WZDrive is a step forward in the development of autonomous systems that can cope with unpredictable conditions. If such models are implemented in vehicles, they may significantly improve road safety, especially where there are infrastructure changes and the presence of people. Research shows that trajectory prediction cannot be based solely on sensor data - understanding human behavior and interaction in the environment is key.

The future of this type of technology depends on further development of datasets, testing in real-world conditions, and collaboration between researchers, vehicle manufacturers, and transportation authorities. WZDrive is not a ready-made solution for mass deployment, but it is an important step towards creating systems that can operate safely even in the most complex road situations. This may be the key to full autonomy in urban environments and construction zones.