Urban traffic monitoring based on aerial LIDAR data

Links:  Video

Related Publications:
ICPR 2012 (Paper),  Attila Börcs and Csaba Benedek
ISPRS 2012 (Paper),   Attila Börcs and Csaba Benedek

Introduction

Automatic traffic monitoring is a central goal of urban traffic control, environmental protection and aerial surveillance applications. We propose a Two-Level MPP (L2 MPP) model, which partitionates the complete vehicle population into vehicle groups, called traffic segments, and extracts the vehicles and the optimal segments simultaneously by a joint energy minimization process. Object interactions are differently defined within the same segment and between two different segments, implementing adaptive object neighborhoods.

Point Cloud Preprocessing

We have developed a Markov Random Field (MRF) model for point cloud segmentation, which utilizes various 3D descriptors. For featuring the terrain class, we estimate the dominant plane of the input cloud using the RANSAC algorithm. Regarding the roof class, we assume that roof points form large connected regions of the cloud, which are composed of segments with uniform surface normals. Local point cloud density is also calculated to recognize sparse clutter regions (like most reflections from walls), and echo number is exploited for identifying vegetation. Points of vehicles appear as outliers from the previous classes, and in addition, they fit prior height distributions of the possible cars. Finally, we construct an MRF energy model based on the previous features. For optimization, the fast ICM algorithm proved to be efficient in the used test sets.






Grid Projection

After the 3D segmentation process, we stretch a 2D pixel lattice S, i.e. an image, onto the ground plane, where s ∈ S denotes a single pixel. Then, we project each LIDAR point to this lattice, which has a label terrain, building roof (purple), ground (blue) or vehicle (black). Since the projection of the sparse point cloud to a regular image lattice may result in pixels without definite class labels and intensities, we also use undefined labels at certain pixels (white pixels Terrain and roof regions are jointly referred as background.






Two-Level MPP (L2 MPP) model

Inverse methods, such as Marked Point Processes, MPPs, assign a fitness value t o each possible object configuration, thereafter an optimization process attempts to find the configuration with the highest confidence. In this way complex object appearance models can be used, it is easy to incorporate prior shape information (e.g. only searching among rectangles) and object interactions (e.g. penalize intersection, favor similar orientation). Conventional MPP models offer limited options for hierarchical scene modeling, since they usually exploit pairwise object interactions, which are defined on fixed symmetric object neighborhoods. In a traffic situation we often find several groups of regularly aligned vehicles, but we must also deal with junctions or skewed parking places next to the roads where many differently oriented cars appear close to each other. In addition, the coherent car groups may have thin, elongated shapes, therefore concentric neighborhoods are less efficient. For this reason, we propose here a Two-Level MPP (L2 MPP) model, which partitionates the complete vehicle population into vehicle groups, called traffic segments, and extracts the vehicles and the optimal segments simultaneously by a joint energy minimization process.