6. Conclusions
The flexible software architecture presented in this paper demonstrates adaptation to different distributed ToF sensor configurations. The system combines a hardware triggering scheme, 3D point cloud registration with a continuous fidelity check, probabilistic occupancy grid mapping, SNN-based object detection, and runtime execution monitoring. Notably, by integrating up to seven cameras, the system scales effortlessly, preserving low average latency (<107 ms acquisition, <40 ms fusion, 29 ms inference, and 39 ms real-time visualization).
The introduced external trigger scheme and the synchronization analysis show significant improvements in interference mitigation by applying only a 10 μs delay between staggered triggers. Through point cloud registration accuracy analysis, we highlight the challenges of maintaining global alignment in closed-loop geometries, observing a notable drift up to 0.68 m in direct versus indirect pairwise sequential transformations, which reveals that future refinements could include registration methods to mitigate indirect error propagation.
The custom event-based SNN inference model demonstrated competitive precision of 65% mAP while benefiting from inherently sparse ToF data, making it well suited for real-time, low-power applications. Further optimizations could include SNN model quantization for neuromorphic hardware acceleration to enhance computational efficiency. While the pretrained networks detected objects in high-density ToF sensor point clouds, the LiDAR-to-ToF domain gap requires careful preprocessing, evidenced by PV-RCNN’s improved performance with double voxelization (52% to 59% improvement). Likewise, retraining models on the non-scanning LiDAR dataset showed improvements due to the closer similarity to ToF data. Future work could boost accuracy by incorporating synthetic ToF sensor data from the CARLA [78] simulator into the training set to better bridge this gap and enhance generalization. In terms of the ToF system technology transfer to real-world automated driving solutions, identifying vulnerable road user (VRU) object categories with ToF sensors and data-driven models could offer unique advantages regarding technology acceptance, primarily due to the ability to capture dense depth data instead of recognizable visual details, removing the privacy concerns related to facial identification. Additionally, prioritizing a safe zone around the vehicle and achieving demonstrable reliability in reducing accidents would gain trust and bolster societal acceptance of AD safety systems.
Lastly, despite the notable limitations we observed with the chosen current generation sensors, such as occasional visible distance reduction in bright sunlight, hardware improvements could address these constraints and thorough characterization against environmental conditions could be conducted. By refining the sensor setup—for example, by incorporating higher range sensors in critical blind spots—the system could be further adapted for specific applications, making the approach adaptable to a wide range of autonomous perception tasks. Moreover, the simplistic low-level software architecture allows for deployment on embedded heterogeneous RISC architectures (AArch64, RISV-V) with neuromorphic NPU integration for low-power operation on resource-constrained platforms."Interestingly this is not the first time that AKIDA technologies compatibility with Time of Flight Sensors has been proven nor is it a novel idea that AKIDA technology might be integrated with RISC architectures: see Frontgrade Gaisler and GRAIN and Si Fives Intelligence Series incorporating AKIDA.
My opinion only DYOR
Fact Finder
(20min delay)