Drone, Cybersecurity, Radar Signal Processing, Machine Learning, Kinematics
Aerospace Engineering, Mechanical Engineering, Computer Science, Communications Engineering, Cybersecurity, Electronics Engineering
Multi-copter drone detection and discrimination from birds and manned aircrafts is critical for ensuring safe drone operations.
Radar cross section (RCS) measurement is only second to visible band images as primary data for commercial drone early warning systems. RCS is unaffected by lighting conditions and is much less affected by dust, fog, snow or rain in comparison with visible band images.
The RCS comprises two elements: a) reflections from rigid, non-moving parts of the body in the reference frame of its own centroid and, b) micro doppler which registers as phase shifts in radar returns produced by the motion of its moving parts. The former has a larger magnitude but, typically, less discrimination power. Radar data processing is computationally expensive and classification latency is a concern due to the rather short time detection of an unauthorized drone leaves for safety-enhancing actions.
To improve detection and classification accuracy of multi-copter drones in low signal-to-noise (SNR) environment and reduce latency, we will add kinematics features to micro doppler augmented RCS features. The kinematics features will be extracted from the RCS itself. Thus, there could be several hundred features that will require a computationally expensive classifier. One potential approach to minimize the computational load is to feed only a subset to the classifier. An optimal subset may depend on regime-defining parameters such as SNR, distance-to-flying-object and metadata.
Machine learning (ML) techniques include regression, SVM, decision tree, neural network and deep learning (DL) amongst others, in (a largely) increasing order of computational complexity. To further reduce latency, we will design a hierarchical ML classifier with computational primitives that get progressively demanding as one moves from the root to the leaf. DL classifiers will be optimized via pruning and quantization rationalization.
Matlab/Python, C++, experience with SolidWorks or AutoCAD is helpful though not necessary
An understanding of computer architecture and hands-on experience with Arduino or Raspberry Pi is helpful though not necessary
India : Dr. Dhanesh G. Kurup, industry
Outside : They will be arranged once you have a substantial conference publication arising out of your PhD work.
Amrita Vishwa Vidyapeetham provides stipends and teaching assistantships to selected candidates. Once you join, you could leverage existing proposals within the team to apply for additional corporate or government funding
3 ½ to 4 ½ years based on full-time committed
Professor,
Electrical & Electronics Engineering,
School of Engineering, Coimbatore