Анализ компонентов и альтернативы

Эта страница фиксирует глубокий component analysis за новым phase split. Текущий hardware полезен, но его нельзя считать final production design.

Summary recommendation

Product phase Recommended hardware posture Why
Phase 1 - Lab validation Keep Jetson Orin Nano + USB UVC camera. Fastest model iteration and easiest Linux video debugging.
Phase 2 - Field MVP Use Jetson for first real-hive pilot unless measured Pi + Hailo prototype is ready. Harden enclosure, power, mounting and network first. Field failures will come from weather, power, cable, camera placement and support gaps before compute cost matters.
Phase 3 - Production kit Select from Jetson production SOM/carrier, Raspberry Pi 5 + Hailo, RK3588 or smart camera based on measured count accuracy per watt. Production must optimize accuracy, energy, bandwidth, serviceability and BOM cost together.

Compute platform analysis

Platform Strengths Weaknesses Best use Recommendation
NVIDIA Jetson Orin Nano Super Developer Kit Strong CUDA/TensorRT ecosystem, high AI headroom, easy PyTorch/Ultralytics path, good Docker workflow. Higher power/cost, developer-kit form factor, thermal care required, weak low-power video encoding story. Lab reference, model development, powered pilot, premium/dev SKU. Keep now; do not assume production default.
Raspberry Pi 5 + Hailo AI HAT+ 26 TOPS Lower cost, likely lower power, strong camera ecosystem, community, dedicated AI accelerator. Requires model conversion and HailoRT pipeline; tracker may still use Pi CPU; thermal needs testing. First cost-down production candidate. Build side-by-side benchmark after Jetson baseline stable.
Raspberry Pi 5 + Hailo AI HAT+ 13 TOPS Cheaper than 26 TOPS and may be enough for tiny model. Less headroom for tracking/future models/overlays. Later cost-down variant. Test only after 26 TOPS path works.
RK3588 NPU board Attractive integrated NPU and media-codec potential. RKNN tooling, board quality, OS support and camera support vary. Media-heavy alternative if Hailo conversion poor or H.265 matters most. Third priority, but worth controlled benchmark.
Coral TPU Very low power for supported TFLite models. Current YOLO-style bee detector may require major simplification. Telemetry-only or tiny-model experiments. Defer until simplified model exists.
Smart/IP camera with onboard AI or encoder Integrated sensor, enclosure and H.265 can simplify video product. Vendor lock-in, limited model control, API uncertainty, frame access latency. Long-term integrated field SKU research. Research after dataset/FOV/count logic stable.

Compute decision metrics

Metric Why it matters Production target direction
Count accuracy per watt Combines user value with energy cost. Choose lowest-energy platform meeting count-quality tolerance.
Full-pipeline FPS Detector-only FPS ignores tracking, counting, upload and watchdog overhead. 15 FPS at 720p or 30 FPS at 640x480 as first target.
FPS/W Converts benchmark results into solar/thermal requirements. Prefer at least 2x improvement before switching from Jetson reference.
Video encode load Live/clips must not starve AI. Hardware H.264/H.265 or camera-side encode preferred.
Thermal stability Enclosures get hot. No throttling that changes count behavior during 4-8 hour soak.
Maintainability Field support matters as much as speed. Stable OS, drivers, logs, remote update and watchdog support.

Camera and optics analysis

Current or candidate Strengths Risks Better direction
MOKOSE 4K USB UVC camera Easy Linux debugging, many resolutions, no CSI driver work, quick replacement. Bulky USB cable, uncertain supply, rolling shutter/exposure unknown, not sealed camera head. Keep for lab and first pilot only if outdoor image tests pass.
Manual varifocal CS/C lens Helps discover field of view, distance and entrance coverage. Installer can change focus/zoom; settings drift; focal length can miss landing board. Use for discovery, then freeze fixed focal length or lock lens mechanically.
CSI/MIPI board camera Compact, lower cabling bulk, production-like integration. FPC cable length and robustness are harder outdoors; drivers must be verified. Evaluate after FOV known and enclosure concept exists.
IP camera with H.265 Camera handles video compression and outdoor housing may be mature. Edge AI needs frame access; latency/vendor APIs can complicate counting/live sessions. Candidate for video-heavy SKU if frame access reliable.
Global-shutter camera Reduces fast-motion distortion. Higher cost and sometimes lower low-light quality. Test only if rolling shutter causes count errors in labelled clips.

Camera-head production requirements

Requirement Why it matters Practical implementation
Fixed geometry Count lines and regions depend on stable view. Bracket with alignment marks and documented distance/angle.
Locked optics Focus drift looks like model regression. Fixed-focus lens or lock screw/thread-lock after setup.
Lighting control Bee entrance lighting changes quickly. Hood, constrained exposure, sample clips at dawn/noon/cloud/rain.
Clean window Glare, scratches, water drops and condensation reduce detection quality. Tilted AR-coated window, hood, drainage path, service cleaning access.
Replaceable head Cameras fail or cables get damaged. Sealed connector or replaceable camera-head subassembly.

Video encoding and bandwidth analysis

Product should not stream continuous video. It should upload movement telemetry continuously and publish high-quality video only when someone watches or when selected clips are needed.

Mode Recommended behavior Reason
Telemetry default Upload movement and health metrics. Small, durable, useful for charts and alerts.
On-demand live Publish media through gate-video-stream only while session active. Avoids constant bandwidth and keeps device private behind outbound connections.
Event clip Upload anomaly/manual/QA samples. Supports review and model improvement without endless storage.
Service visit Allow local high-quality clip download. Avoids cellular data costs for dataset collection.
Continuous recording Lab only. Too expensive/noisy as field default.

Storage analysis

Storage option Use Risk Recommendation
Consumer NVMe SSD Lab and early field buffering. Consumer storage may dislike heat, power loss or vibration. Good for Phase 1-2; add retention limits and disk health telemetry.
Industrial microSD/eMMC Low-cost production logs and small clips. Lower capacity/write endurance if clip policy too heavy. Use only with strict retention/write-rate limits.
Industrial NVMe/eMMC Video-heavy production SKU. Higher cost. Use when local video retention is paid feature.
Cloud-only video No local storage beyond queue. Fails during network outages. Avoid for field units; keep local queue/buffer.

Network analysis

Option Strengths Weaknesses Best fit
Ethernet/PoE Reliable data and power in one cable. Requires infrastructure near hive. First customer-grade powered installations.
WiFi Cheap/easy for hobby apiaries near buildings. Weak signal, enclosure RF issues, retries needed. Field MVP and urban SKU.
LTE router/modem Works in remote apiaries. Data plan, current bursts, antenna placement, recurring cost. Remote paid SKU or shared apiary gateway.
Local gateway One backhaul for multiple hives. More system complexity. Apiaries with many devices.
LoRa telemetry Very low power and long range. Not suitable for live video. Telemetry-only bee-count summaries.

Power analysis

Power design Strengths Risks Recommendation
Bench USB-C/DC Fast development. Not field-safe by itself. Phase 1 only.
Outdoor mains to low-voltage DC Stable pilot power. Must avoid unsafe mains inside hobby enclosure. Good Phase 2 pilot path.
PoE Solves power and data together. Needs Ethernet infrastructure. Best powered production path where available.
Battery-only Good for short energy experiments. Runtime may be short for all-day vision. Use to measure Wh/day, not default product claim.
Solar + battery Attractive for remote apiaries. Large/expensive with Jetson-class active power. Production SKU only after measured duty-cycle and low-power modes.

Enclosure and mechanical analysis

Current item Assessment Production alternative
2020 aluminium extrusion Great for lab fixtures and adjustable mounts. Custom/repeatable bracket with corrosion-resistant fasteners.
Acrylic/plexiglass sheets Useful for glare/condensation experiments. UV-stable, cleanable, tilted optical window or camera hood.
Acrylic Jetson case Not weatherproof and may not fit/cool Orin. IP-rated electronics box with thermal/cable-entry design.
Adjustable 1/4 camera mount Good for discovery/alignment tests. Locked bracket with alignment marks and service instructions.
Loose USB/HDMI cables Fine on desk. Sealed cable entries, strain relief, no permanent display cable.

Final selection rule

Production component is acceptable only if it passes four checks:

  1. Data quality — count accuracy and confidence are good enough on labelled entrance clips.
  2. Energy and thermal — platform runs target duty cycle without overheating or oversizing power.
  3. Serviceability — support can diagnose camera, model, network, disk and power state remotely.
  4. Manufacturability — stable supply, documented assembly, replaceable harnesses and repeatable QA tests.