Chen-Fu Yeh
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Research Interests
Chen-Fu Yeh has a total of 15 co-authors. Here are the top co-authors by number of shared publications:
- Alexander James White
Institute of Systems Neuroscience, National Tsing Hua University, Hsinchu, Taiwan
1 shared publication Chao-Yang Tang
1 shared publicationCheng-Te Wang
Institute of Systems Neuroscience, National Tsing Hua University, Hsinchu, Taiwan
1 shared publicationChih-Cheng Hsieh
1 shared publication- Chih-Cheng Hsieh
National Tsing Hua University
1 shared publication Chung-Chuan Lo
1 shared publication
Latest Publications
- Open Access
- 3 April 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in TechRxiv
FlowDep - An efficient and optical-flow-based algorithm of obstacle detection for autonomous mini-vehicles
Obstacle detection is crucial for the safety and efficiency of autonomous vehicles. For mini-vehicles such as palm-sized drones, it is a challenge to implement traditional methods like Lidar due to high costs and physical constraints. Vision-based deep learning approaches, while accurate, are too resource-intensive for the mini-vehicles. To address this issue, we introduce Flowdep, a novel optical-flow-based algorithm inspired by the low-resolution but efficient motion-detection mechanisms in insects. Flowdep combines optic flow and IMU (or positioning information) to estimate the depth of every image pixel. We also generate a variant of Flowdep using the artificial neural network (Flowdep-ANN). Our tests show that Flowdep and Flowdep-ANN are 5.8 to 114.7 times faster than the DNN networks we tested, while the accuracies of Flowdep and Flowdep-ANN are on par with these networks. We further tested Flowdep and Flowdep-ANN on a small autonomous vehicle with Raspberry Pi4 as the computing platform, and both models successfully performed real-time object detection. The present work demonstrates the potential of using optical flow as an efficient approach to estimate depth and detect obstacles in resource-constrained mini-vehicles.
- 3 April 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in TechRxiv
Open AccessFlowDep - An efficient and optical-flow-based algorithm of obstacle detection for autonomous mini-vehicles
Obstacle detection is crucial for the safety and efficiency of autonomous vehicles. For mini-vehicles such as palm-sized drones, it is a challenge to implement traditional methods like Lidar due to high costs and physical constraints. Vision-based deep learning approaches, while accurate, are too resource-intensive for the mini-vehicles. To address this issue, we introduce Flowdep, a novel optical-flow-based algorithm inspired by the low-resolution but efficient motion-detection mechanisms in insects. Flowdep combines optic flow and IMU (or positioning information) to estimate the depth of every image pixel. We also generate a variant of Flowdep using the artificial neural network (Flowdep-ANN). Our tests show that Flowdep and Flowdep-ANN are 5.8 to 114.7 times faster than the DNN networks we tested, while the accuracies of Flowdep and Flowdep-ANN are on par with these networks. We further tested Flowdep and Flowdep-ANN on a small autonomous vehicle with Raspberry Pi4 as the computing platform, and both models successfully performed real-time object detection. The present work demonstrates the potential of using optical flow as an efficient approach to estimate depth and detect obstacles in resource-constrained mini-vehicles.
- 6 June 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-4
Integer Quadratic Integrate-and-Fire (IQIF): A Neuron Model for Digital Neuromorphic Systems
Simulation of a spiking neural network involves solving a large number of differential equations. This is a challenge even for modern computer systems, especially when simulating large-scale neural networks. To address this challenge, we design a neuron model: the Integer Quadratic Integrate-and-Fire (IQIF) neuron. Instead of computing on floating point numbers, as is typical with other spiking neuron models, the IQIF model is computed purely on integers. The IQIF model is a quantized and linearized version of the classic quadratic integrate-and-fire (QIF) model. The IQIF model retains all dynamic characteristics of the QIF model with much lower computation complexity, at the cost of a limited dynamic range of the membrane potential and the synaptic current. We compare IQIF to other spiking neuron models based on their simulation speeds and the number of neuronal behaviors they can perform. We further compare the performance of IQIF with the leaky integrate-and-fire model in a classical decision-making network that exhibits nonlinear attractor dynamics. Our results show that the IQIF neurons are capable of performing computation that other spiking neuron models can do while having the advantages of speed. Moreover, the IQIF model is digital hardware friendly due to its pure integer operation and is therefore easily to be implemented in custom-built neuromorphic systems.
- 6 June 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-4
Integer Quadratic Integrate-and-Fire (IQIF): A Neuron Model for Digital Neuromorphic Systems
Simulation of a spiking neural network involves solving a large number of differential equations. This is a challenge even for modern computer systems, especially when simulating large-scale neural networks. To address this challenge, we design a neuron model: the Integer Quadratic Integrate-and-Fire (IQIF) neuron. Instead of computing on floating point numbers, as is typical with other spiking neuron models, the IQIF model is computed purely on integers. The IQIF model is a quantized and linearized version of the classic quadratic integrate-and-fire (QIF) model. The IQIF model retains all dynamic characteristics of the QIF model with much lower computation complexity, at the cost of a limited dynamic range of the membrane potential and the synaptic current. We compare IQIF to other spiking neuron models based on their simulation speeds and the number of neuronal behaviors they can perform. We further compare the performance of IQIF with the leaky integrate-and-fire model in a classical decision-making network that exhibits nonlinear attractor dynamics. Our results show that the IQIF neurons are capable of performing computation that other spiking neuron models can do while having the advantages of speed. Moreover, the IQIF model is digital hardware friendly due to its pure integer operation and is therefore easily to be implemented in custom-built neuromorphic systems.