Probably Unknown: Deep Inverse Sensor Modelling In Radar (1810.08151v2)

Abstract: Radar presents a promising alternative to lidar and vision in autonomous vehicle applications, able to detect objects at long range under a variety of weather conditions. However, distinguishing between occupied and free space from raw radar power returns is challenging due to complex interactions between sensor noise and occlusion. To counter this we propose to learn an Inverse Sensor Model (ISM) converting a raw radar scan to a grid map of occupancy probabilities using a deep neural network. Our network is self-supervised using partial occupancy labels generated by lidar, allowing a robot to learn about world occupancy from past experience without human supervision. We evaluate our approach on five hours of data recorded in a dynamic urban environment. By accounting for the scene context of each grid cell our model is able to successfully segment the world into occupied and free space, outperforming standard CFAR filtering approaches. Additionally by incorporating heteroscedastic uncertainty into our model formulation, we are able to quantify the variance in the uncertainty throughout the sensor observation. Through this mechanism we are able to successfully identify regions of space that are likely to be occluded.

• The paper introduces a deep inverse sensor model that efficiently converts radar data into occupancy probabilities using neural networks.
• The paper employs self-supervised learning with partial lidar labels to overcome sensor noise and occlusion in urban driving scenarios.
• The paper integrates heteroscedastic uncertainty modeling, enhancing obstacle detection and scene segmentation in challenging conditions.

Overview of "Probably Unknown: Deep Inverse Sensor Modelling In Radar"

The paper "Probably Unknown: Deep Inverse Sensor Modelling In Radar" presents a novel approach for processing radar data in autonomous vehicle applications by leveraging deep learning techniques to enhance scene segmentation into occupied and free spaces. Radar, an advantageous sensor modality over lidar in adverse weather conditions, poses challenges in interpreting raw data due to sensor noise and occlusion. This paper addresses these challenges by introducing a deep inverse sensor model (ISM) that efficiently segments radar data into occupancy grids using a neural network.

Key Contributions

The paper outlines several contributions, including:

1. Deep Inverse Sensor Model: A neural network-based ISM is introduced to convert radar data into occupancy probabilities. Unlike classical filtering approaches such as Constant False-Alarm Rate (CFAR), this method utilizes a deep learning framework to account for scene context, thus outperforming traditional techniques.
2. Self-Supervised Learning: The ISM is trained using self-supervision with partial labels generated from lidar data, eliminating the need for manual annotations while enabling continuous learning from environmental interactions.
3. Handling Uncertainty: By incorporating heteroscedastic uncertainty into the neural network, the paper quantifies varying uncertainties across the sensor observations. This allows the identification of occluded regions, enhancing the model's capability to distinguish between areas of true occupancy and those affected by sensor noise or occlusion.
4. Experimental Validation: The approach is validated on five hours of urban driving data, showing improved performance in Intersection over Union (IoU) scores compared to CFAR methods.

Implications

The implications of this research are significant within autonomous transportation systems. The paper demonstrates that deep learning can dramatically improve radar data interpretation by considering scene context rather than traditional filtering methods reliant on predetermined parameters. This can lead to more reliable navigation and obstacle detection under challenging conditions where lidar might fail.

Future Directions

The authors suggest potential future directions, such as integrating dynamic scene understanding into the ISM framework. This could enable autonomous systems to not only recognize static obstacles but also predict and adapt to moving entities, further enhancing navigational capabilities.

The approach outlined in the paper exemplifies the burgeoning role of AI and deep learning in sensor data processing, marking a shift towards more context-aware and adaptive systems in autonomous vehicles. Further research could explore real-time processing optimizations and broader environmental testing to refine and scale this methodology.

Overall, this paper contributes valuable insights into radar data processing, with implications extending to the advancement of autonomous vehicle technology and intelligent sensor applications.