Fully Convolutional Neural Networks for Dynamic Object Detection in Grid Maps (Masters Thesis)

TL;DR

This thesis enhances dynamic obstacle detection in grid maps using Fully Convolutional Neural Networks, improving classification accuracy and obstacle orientation estimation, with a 34.8% performance gain over previous methods.

Contribution

It introduces FCNN-based classification and orientation estimation for dynamic obstacles, and evaluates semi-supervised learning for data augmentation in grid map analysis.

Findings

34.8% improvement over previous approach

FCNN provides better orientation estimation

Semi-supervised learning did not improve performance

Abstract

One of the most important parts of environment perception is the detection of obstacles in the surrounding of the vehicle. To achieve that, several sensors like radars, LiDARs and cameras are installed in autonomous vehicles. The produced sensor data is fused to a general representation of the surrounding. In this thesis the dynamic occupancy grid map approach of Nuss et al. is used while three goals are achieved. First, the approach of Nuss et al. to distinguish between moving and non-moving obstacles is improved by using Fully Convolutional Neural Networks to create a class prediction for each grid cell. For this purpose, the network is initialized with public pre-trained network models and the training is executed with a semi-automatic generated dataset. The second goal is to provide orientation information for each detected moving obstacle. This could improve tracking algorithms, which are based on the dynamic occupancy grid map. The orientation extraction based on the Convolutional Neural Network shows a better performance in comparison to an orientation extraction directly over the velocity information of the dynamic occupancy grid map. A general problem of developing machine learning approaches like Neural Networks is the number of labeled data, which can always be increased. For this reason, the last goal is to evaluate a semi-supervised learning algorithm, to generate automatically more labeled data. The result of this evaluation shows that the automated labeled data does not improve the performance of the Convolutional Neural Network. All in all, the best results are combined to compare the detection against the approach of Nuss et al. [36] and a relative improvement of 34.8% is reached.