Cooperative Probabilistic Trajectory Forecasting under Occlusion (2312.03296v1)

Abstract: Perception and planning under occlusion is essential for safety-critical tasks. Occlusion-aware planning often requires communicating the information of the occluded object to the ego agent for safe navigation. However, communicating rich sensor information under adverse conditions during communication loss and limited bandwidth may not be always feasible. Further, in GPS denied environments and indoor navigation, localizing and sharing of occluded objects can be challenging. To overcome this, relative pose estimation between connected agents sharing a common field of view can be a computationally effective way of communicating information about surrounding objects. In this paper, we design an end-to-end network that cooperatively estimates the current states of occluded pedestrian in the reference frame of ego agent and then predicts the trajectory with safety guarantees. Experimentally, we show that the uncertainty-aware trajectory prediction of occluded pedestrian by the ego agent is almost similar to the ground truth trajectory assuming no occlusion. The current research holds promise for uncertainty-aware navigation among multiple connected agents under occlusion.

• The paper introduces an innovative method that integrates relative pose estimation with probabilistic forecasting to address occluded trajectories.
• It employs deep probabilistic learning techniques, including Monte Carlo dropout, to quantify uncertainty and outperform traditional Kalman filters.
• Experimental results demonstrate robust performance in various occlusion scenarios, suggesting significant improvements in autonomous navigation safety.

Introduction to Cooperative Probabilistic Trajectory Forecasting

When navigating environments where vision can be obstructed, reliable perception and forecasting of other agents' movements become crucial for safety-critical tasks. Traditional approaches rely on continuous detection, but obstacles can often obscure objects, complicating detection and forecasting efforts. Advancements in communication among connected agents offer new solutions for such occlusions. This paper explores an innovative method to address these challenges by combining relative pose estimation with probabilistic trajectory forecasting.

Embedded Challenges in Trajectory Prediction

Current State of Technology

Accurate object detection and intention forecasting have evolved with end-to-end pipelines converting raw sensor data into predictions. But occlusions still pose a problem, leading to potential inaccuracies and safety concerns.

Cooperative Perception and Communication Limits

While connected vehicles and infrastructure can share detailed sensor data, the process can be bandwidth-intensive and induce latency. A more efficient strategy involves sharing condensed yet crucial information among agents to maintain awareness without overburdening communication channels.

Methodology for Relative Pose Estimation

Feature Detection and Matching

For relative pose estimation between connected agents, key visual features must first be detected and matched across different perspectives. Efficient methods such as the ORB descriptor facilitate this process, which is followed by calculating the relative orientation through transformations derived from these feature correspondences.

Issues with Existing Approaches

Previous research has addressed cooperative perception and orientation estimation. Still, there is limited exploration of the application of these techniques to occlusion-aware prediction.

Advancements in Trajectory Forecasting

Conventional vs. Probabilistic Prediction

While standard models such as Kalman Filters have been used for predictions, occlusions introduce uncertainties that these models are not equipped to handle. The shift towards deep probabilistic learning methods enables models to provide distributions over future states rather than deterministic point estimates, thus enhancing robustness and safety.

Proposed Solution

The paper outlines a novel end-to-end architecture that integrates relative pose recovery with probabilistic trajectory forecasting, aimed at predicting an object's future states under occlusion. A novel aspect is the use of approximate Bayesian inference methods, such as Monte Carlo dropout, allowing for uncertainty quantification in the predicted states.

Achievements and Future Directions

Robustness to Occlusion and Practical Implications

The experiments show the strength of the proposed methodology across multiple scenarios, including partial and intermittent occlusions. Such capabilities hold significant potential for improving navigation and planning in cluttered environments involving multiple connected agents.

Limitations and Future Work

One limitation mentioned is the reliance on static sensors for the experiments. Future research could include integrating dynamic relative pose estimation into the proposed framework, which would broaden the applicability of these findings to more dynamic environments and contribute to the advancement of cooperative perception.

In conclusion, this research offers a forward-thinking approach to cooperative probabilistic trajectory forecasting under occlusion. The method provides a pathway towards more reliable and safe navigation in scenarios where direct perception is hindered, presenting a significant step forward in the domain of autonomous systems and connected vehicles.