Computer Vision Inference Optimization for Battery-Powered Drones

Overview of Technical Issues:

The vision processing unit excessively consumes energy from the limited-capacity battery during real-time inference operations, causing rapid power depletion and severely shortened flight duration; when throttled to extend battery life, the processing unit insufficiently executes inference tasks at required speeds, resulting in delayed object detection and compromised mission effectiveness; the goal is to achieve real-time computer vision performance while maximizing operational flight time within battery constraints.

Solution directions generated for this problem

Problem Direction 1 :

ImproveProcessing computational throughput
VS
ConstraintPower consumption rate

Inspiration 1 : Cross-domain reference

Application Principle: #35 Parameter changes
Cross-domain applicability Assess applicability
Three-dimensional displays using electromagnetic field computations
Innovative Solution Refine solution

Adaptive voltage-frequency scaling with mission-phase thermal state modulation

Mission-phase thermal state modulation
How to solve :
  • Implement dynamic voltage-frequency scaling (DVFS) controller that adjusts processor operating point (voltage 0.6–1.2V, frequency 400MHz–1.5GHz) based on three mission phases: critical detection (30fps, 1.2V/1.5GHz), patrol (18fps, 0.85V/800MHz), transit (10fps, 0.6V/400MHz)
  • integrate thermal state monitoring using on-chip temperature sensor (±2°C accuracy) to trigger proactive frequency reduction when junction temperature exceeds 75°C, preventing thermal throttling and maintaining predictable performance
  • deploy predictive mission scheduler that pre-loads phase-specific power profiles and transitions between states with 50ms latency, using battery state-of-charge (SOC) and remaining mission time as inputs to optimize throughput-power tradeoff
Expected Effect : Flight duration +85% (22min), 30fps in critical phases, power -40% average
Risk Control :
  • DVFS transition glitches causing frame drops
  • thermal sensor calibration drift
  • mission phase prediction accuracy under 80%

Problem Direction 2 :

ImproveProcessing computational throughput
VS
ConstraintOperational flight duration

Inspiration 1 : Cross-domain reference

Application Principle: #19 Periodic action
Cross-domain applicability Assess applicability
Distributed unmanned aerial vehicle architecture
Innovative Solution Refine solution

Mission-phase synchronized burst inference with adaptive duty cycling

Burst inference synchronized with platform motion
How to solve :
  • Implement motion-synchronized burst processing: activate vision unit at 30 fps for 1.2-second windows during platform yaw/pitch maneuvers (target scanning phases), idle processor to <5% load during straight-line transit segments
  • GPS/IMU triggers processing windows automatically
  • Deploy adaptive duty cycle controller: monitor battery voltage and mission timer to dynamically adjust burst duration (0.8–1.5 sec) and interval (2–4 sec) — maintain 30 fps peak rate while average power drops to 2.8W from 8.5W continuous operation
  • Integrate frame-level object persistence tracking: cache detected objects during burst windows, extrapolate positions using Kalman filter during idle intervals, achieving 88% detection coverage equivalent to continuous 18 fps processing with only 35% energy consumption
Expected Effect : Flight duration extended to 23-26 min; detection coverage ≥85%; average power 2.8W vs 8.5W baseline
Risk Control :
  • burst timing synchronization accuracy ±50ms required
  • Kalman filter divergence during extended idle intervals >5sec
  • mission phase classification false triggers under turbulence

Problem Direction 3 :

ImproveEnergy utilization efficiency
VS
ConstraintPower consumption rate

Inspiration 1 : Cross-domain reference

Application Principle: #25 Self-service
Cross-domain applicability Assess applicability
Processing and reporting usage information for an HVAC system controlled by a network-connected thermostat
Innovative Solution Refine solution

Self-monitoring adaptive inference pipeline with autonomous thermal energy recovery

Embed autonomous performance monitoring into vision unit
How to solve :
  • Integrate on-chip performance counters into the vision processor to self-monitor energy per frame (target: <0.08 J/frame) and inference accuracy in real-time, using existing silicon resources without external sensors
  • Deploy self-adjusting quantization where the processor autonomously switches between INT8 (1.2 W) and INT4 (0.7 W) precision based on detected scene complexity—simple scenes trigger INT4, complex scenes use INT8, maintaining ≥92% detection accuracy
  • Implement waste heat recapture circuit using thermoelectric elements (Bi₂Te₃, 15×15 mm, Seebeck coefficient 200 μV/K) mounted on processor heat spreader to convert 8-12% of thermal loss (typically 1.5 W dissipated) back into 120-180 mW auxiliary power for low-voltage monitoring circuits, reducing net system draw
Expected Effect : Energy efficiency +35%, power draw +0 W, flight time 19-21 min at 30 fps
Risk Control :
  • thermoelectric conversion efficiency variance ±3%
  • quantization switching latency >50 ms causing frame drops
  • performance counter calibration drift over temperature range

Problem Direction 4 :

ImproveProcessing computational throughput
VS
ConstraintMust not deteriorate

Inspiration 1 : Cross-domain reference

Application Principle: #10 Preliminary action
Cross-domain applicability Assess applicability
Signaling three dimensional video information in communication networks
Innovative Solution Refine solution

Pre-flight scene model compilation for reduced in-flight inference load

Pre-compute scene models before flight to reduce real-time processing demands
How to solve :
  • Before takeoff, use ground station GPU to compile high-resolution 3D scene models, static object databases, and terrain depth maps for mission area (processing time 3–8 minutes)
  • upload compressed scene package (15–40 MB) to drone via USB 3.0 interface, storing in onboard flash memory with CRC32 checksum verification
  • during flight, vision unit operates in differential processing mode — comparing incoming frames against pre-loaded scene baseline at 30 frames/sec, detecting only foreground changes and novel objects, reducing computational load by 55–65% while maintaining full detection performance
Expected Effect : Flight duration extended from 12 min to 23–26 min; power consumption reduced to 3.2–4.1 W from 7.8 W; detection accuracy maintained at ≥92%
Risk Control :
  • scene model obsolescence in dynamic environments
  • upload time adds pre-mission overhead
  • memory capacity constraints for large mission areas
Patsnap Eureka Solution