Abstract
In this paper, we present an approach for the detection of persons in point clouds gathered by mobile laser scanning (MLS) systems. The approach consists of a preprocessing stage and the actual detection. The main task of the preprocessing is to reduce the amount of data which has to be processed by the detection. To fulfill this task, the preprocessing consists of ground removal, segmentation, and several filters. The detection is based on an implicit shape models (ISM) approach, which is an extension to bag-of-words approaches. For this detection method, it is sufficient to work with a small amount of training data. Although in this paper we focus on the detection of persons, our approach is able to detect multiple classes of objects in point clouds. Using a parameterization of the approach which offers a good compromise between detection and runtime performance, we are able to achieve a precision of 0.68 and a recall of 0.76 while having an average runtime of 370 ms per single scan rotation of the rotating head of a typical MLS sensor.
Introduction
Detecting pedestrians in 3D point clouds is a significant challenge for autonomous systems. In a typical scan from a sensor like the Velodyne HDL-64E, a person might be represented by only 100 to 200 points out of over 130,000. This "needle in a haystack" problem is compounded by the sparsity of data at increasing distances from the sensor, which makes consistent identification difficult.
Current deep learning approaches often require massive, labeled datasets and significant computational power. Autonomous vehicles and robots need to process this data at the speed of acquisition to ensure safety. This research aims to bridge the gap between high detection accuracy and the strict runtime constraints of mobile systems, all while working with limited training examples.
Methodology
The proposed system architecture is divided into two primary phases: Preprocessing and Person Detection.
1. Preprocessing (The Data Sieve)
To achieve real-time performance, the system first discards irrelevant data. This "data sieve" reduces the average point count from approximately 98,000 to just 2,200 per scan—a 98% reduction that allows the detection algorithm to run efficiently.
- Ground Removal: A grid-based approach with recursive traversal identifies and removes the road surface.
- Segmentation: Remaining points are grouped into clusters using Euclidean distance.
- Filtering: Clusters are evaluated based on size and aspect ratio; those not matching a human profile (e.g., long walls or tiny noise clusters) are removed.
2. Person Detection via Implicit Shape Models (ISM)
Instead of holistic detection, the ISM approach identifies local "parts" and their spatial relationship to an object's center. This makes the system robust to partial occlusions.
- Feature Description: "Spin images" (2D histograms of point distribution) are calculated for every point in a segment to describe local geometry.
- Dictionary Matching: These descriptors are matched against a pre-trained dictionary of "geometric words."
- Spatial Voting: Each matched word casts a vote for the potential center of a person. A Gaussian weighting function ($\sigma$) scores these positions, allowing for spatial uncertainty.
- Refinement: Potential positions are rated, filtered by score, and merged if they fall within a certain distance of each other.
Experimental Results
The system was evaluated using data from the MODISSA measurement vehicle. The training set was intentionally small, consisting of only 244 person segments and 387 non-person segments.
Key findings include:
- Detection Performance: The system achieved a precision of 0.68 and a recall of 0.76.
- Computational Efficiency: The total average runtime was optimized to 370 ms per scan rotation.
- Search Optimization: A critical insight was that reducing the precision of the dictionary search (from 95% to 50%) had negligible impact on detection performance but reduced the search runtime from 252ms to 32ms.
Conclusion and Future Work
The ISM-based approach demonstrates that effective person detection is possible in sparse MLS data without massive training sets. Future work will focus on parallelizing the segmentation step, integrating temporal tracking to improve reliability, and exploring detection in unsegmented clouds to mitigate errors where pedestrians are merged with background structures.
Summary updated