Vehicle Detection from 3D Lidar Using Fully Convolutional Network (1608.07916v1)

Abstract: Convolutional network techniques have recently achieved great success in vision based detection tasks. This paper introduces the recent development of our research on transplanting the fully convolutional network technique to the detection tasks on 3D range scan data. Specifically, the scenario is set as the vehicle detection task from the range data of Velodyne 64E lidar. We proposes to present the data in a 2D point map and use a single 2D end-to-end fully convolutional network to predict the objectness confidence and the bounding boxes simultaneously. By carefully design the bounding box encoding, it is able to predict full 3D bounding boxes even using a 2D convolutional network. Experiments on the KITTI dataset shows the state-of-the-art performance of the proposed method.

• The paper demonstrates an innovative method using a fully convolutional network to predict complete 3D bounding boxes from transformed 2D Lidar point maps.
• It achieves state-of-the-art performance on the KITTI dataset by efficiently detecting vehicles and accurately estimating their orientations.
• The approach offers enhanced computational efficiency and robustness, paving the way for improved autonomous driving perception systems.

Overview of "Vehicle Detection from 3D Lidar Using Fully Convolutional Network"

The paper "Vehicle Detection from 3D Lidar Using Fully Convolutional Network" by Bo Li, Tianlei Zhang, and Tian Xia presents an innovative method for detecting vehicles using 3D Lidar data. The authors leverage the capabilities of Fully Convolutional Networks (FCNs) to perform object detection on 3D range scan data, specifically focusing on the Velodyne 64E Lidar sensor. This method is particularly relevant for autonomous driving systems that require accurate localization of obstacles.

Methodology

The proposed approach involves transforming 3D Lidar point cloud data into a 2D point map representation. This transformation enables the use of a 2D FCN to simultaneously predict objectness confidence scores and 3D bounding boxes. By cleverly designing the bounding box encoding, the authors are able to predict full 3D bounding boxes despite using a 2D network architecture. Notably, this transformation facilitates computational efficiency and simplifies the application of established 2D CNN techniques to 3D data.

The architecture of the network is described in detail, involving several convolutional and deconvolutional layers that perform down-sampling and up-sampling to extract and categorize features from the Lidar data. The architecture splits into two branches: one for objectness classification and another for 3D bounding box regression.

Results and Performance

The authors validate their approach on the KITTI dataset, a notable benchmark in autonomous driving research. The results exhibit state-of-the-art performance in vehicle detection tasks, measured in terms of Average Precision (AP) and Average Orientation Similarity (AOS). The framework shows enhanced precision in detecting vehicles within a close range, which is crucial for autonomous navigation and obstacle avoidance.

The transformation method used for bounding box encoding ensures that the bounding box predictions remain invariant to the vehicle's position relative to the sensor. This attribute is particularly advantageous in scenarios where the same vehicle may appear differently based on its pose and orientation.

Comparison with Existing Methods

The study compares the proposed method with traditional Lidar-based object detection methods. In particular, it highlights the improved detection capabilities of the FCN approach over methods that rely on handcrafted features. The ability of the FCN to learn comprehensive feature representations contributes significantly to its competitive performance.

Implications and Future Work

This research has substantial implications for advancing autonomous vehicle technology. By integrating FCNs with Lidar data for perception tasks, autonomous systems can achieve a higher level of accuracy and reliability in real-world environments. The potential for expanding this method to other types of Lidar devices and even to RGBD images is also noteworthy.

The paper suggests that further improvements could be achieved through the accumulation of more training data and the design of deeper networks. Such future developments could enhance the model's robustness and generalization capabilities, essential factors for deployment in diverse environmental conditions.

Conclusion

In conclusion, this work represents a significant step forward in utilizing deep learning for Lidar-based vehicle detection. By exploiting the strengths of FCNs and adapting them to 3D data, the paper demonstrates an effective way to harness convolutional networks beyond traditional image processing tasks. As the field of autonomous driving continues to evolve, such innovations will be critical in overcoming the challenges associated with real-time object detection and navigation.