Evaluating the Benefits of Spatio-Temporal Relational Operations for Validating LiDAR Perception Systems
| dc.contributor.author | Holmgren, Jakob | |
| dc.contributor.author | Nord, Philip | |
| dc.contributor.department | Chalmers tekniska högskola / Institutionen för data och informationsteknik | sv |
| dc.contributor.examiner | Sands, David | |
| dc.contributor.supervisor | Massimiliano Gulisano, Vincenzo | |
| dc.date.accessioned | 2021-09-10T07:23:01Z | |
| dc.date.available | 2021-09-10T07:23:01Z | |
| dc.date.issued | 2021 | sv |
| dc.date.submitted | 2020 | |
| dc.description.abstract | On the road to fully autonomous vehicles, advancements in sensor technology haveenabled Advanced Driver-Assistance Systems (ADAS) in assisting the driver with tasks such as lane-keeping and collision avoidance. ADAS systems are enabled by perception systems which, given sensor input, outputs a perception model. Recently proposed perception models utilize data from the Light Detection And Ranging (LiDAR) sensor, that measures nearby occlusions with high definition and frequency. This enables perception models to detect objects and identify environmental segments relevant to ADAS tasks.The critical safety requirements of perception models have created a demand for a framework that can efficiently assess the performance of the model on large amounts of labeled sensor data. Scalable analysis frameworks such as Apache Spark allow developers to perform large-scale data analysis using abstract interfaces while exploiting data-level parallelism in a cluster of nodes. As Spark is a general platform designed for business intelligence and machine learning, there is a trade-off in both expressive power and efficiency for specialized data types. This trade-off is relevant for processing LiDAR data and by extension validating perception models. This lack of specialization impacts model validation such that only expert users have access to efficient implementations. Non-expert users are forced to write complex custom functions for their data which impact efficiency negatively. This increases the risk of errors in the implementation as custom functions are more difficult to interpret. As the query compiler of Spark is inherently extensible, we investigate how its query language, optimizer, and engine have been extended with support for specialized operations on spatial and spatio-temporal data. Furthermore, we investigate how these extensions perform when applied to validating LiDAR sensor perception systems. In our pre-study, we identified spatial joins between volumes and point to be useful unsupported data operations. By implementing these operations with different approaches in Spark, we evaluate the expressive power and efficiency they provide. Finally, we perform a scalability analysis of the query response time of each implementation. We find that using spatial libraries does not improve the efficiency for the specific spatio-temporal join. Furthermore, we find that implementing the query and predicate is not easier with the specialized libraries. However, we find that accounting for different types of shapes of the volumes may impact efficiency on larger datasets. | sv |
| dc.identifier.uri | https://hdl.handle.net/20.500.12380/304103 | |
| dc.language.iso | eng | sv |
| dc.setspec.uppsok | Technology | |
| dc.subject | LiDAR | sv |
| dc.subject | Apache Spark | sv |
| dc.subject | Spatial | sv |
| dc.subject | Spatiotemporal | sv |
| dc.subject | 3D | sv |
| dc.subject | Point Cloud | sv |
| dc.title | Evaluating the Benefits of Spatio-Temporal Relational Operations for Validating LiDAR Perception Systems | sv |
| dc.type.degree | Examensarbete för masterexamen | sv |
| dc.type.uppsok | H |