Computer Vision for Bee Counting and Hive Activity Monitoring
How camera-based systems detect, track and count bees at the hive entrance to build activity indices, and the practical model and hardware choices behind them.
Why entrance activity is a useful proxy signal
A camera aimed at the hive entrance captures a continuous stream of foraging traffic that, when reduced to simple in-and-out counts, functions as a useful proxy for colony health and foraging conditions without ever opening the hive. Activity typically rises through the morning as temperatures climb, peaks during good weather in the main forage season, and drops sharply in poor weather, cold snaps, or when a colony is queenless or diseased and has fewer active foragers to send out. None of this replaces direct inspection, but a sudden, unexplained drop in entrance traffic relative to a colony's own recent baseline, or relative to neighbouring hives experiencing the same weather, is a useful early flag worth investigating.
The engineering challenge is that a busy entrance in peak season can see hundreds of bees passing per minute in both directions, overlapping and occluding one another, which is far harder for a vision system to handle reliably than counting sparse, well-separated objects.
The detection-tracking-counting pipeline
Most working systems follow a similar three-stage pipeline. First, an object detection model, commonly a lightweight variant of YOLO or similar single-shot detectors optimised to run on small edge devices, locates individual bees in each video frame as bounding boxes. Second, a tracking algorithm such as BYTE or OC-SORT links detections across consecutive frames into consistent tracks for each bee, which matters because a single detection frame cannot tell direction of travel on its own; a track that crosses a virtual line at the entrance moving outward counts as an exit, and one crossing inward counts as an entry. Third, the counted crossings are aggregated into per-minute bins and typically smoothed with an exponentially weighted moving average or median filter to reduce noise from momentary detection errors, since raw per-frame counts are jumpy in ways that don't reflect real colony behaviour changes.
Building this into a usable activity index generally requires normalising against each hive's own baseline rather than using a single absolute count across all hives, since entrance size, camera angle and typical colony strength vary enough between hives that a raw count comparison across colonies is often misleading.
Hardware trade-offs: where the processing happens
Running the full detection and tracking pipeline requires meaningful compute, and beekeeping camera systems generally split into edge-processing designs, where a small onboard computer such as a Raspberry Pi or Jetson board runs the model locally and only sends aggregated counts to the cloud, versus cloud-processing designs that stream raw or lightly compressed video to a remote server for analysis. Edge processing suits remote apiaries with weak connectivity since it needs to transmit only a tiny summarised data stream, but it constrains the model to whatever a small board can run in real time, which typically means simpler, faster models rather than the largest and most accurate ones.
Cloud processing can run larger, more accurate models but demands enough bandwidth to move video data continuously, which is often the harder constraint at a rural apiary than the compute itself. Many practical systems use a hybrid approach: capturing continuously at the edge but only uploading short clips triggered by an anomaly, balancing bandwidth against responsiveness.
Frequently Asked Questions
Can a camera-based counting system tell direction, not just total activity?
Yes, this is standard in modern systems. By tracking individual bees across frames and checking which way they cross a virtual line at the entrance, the system distinguishes bees leaving to forage from bees returning, which is far more informative than a raw total traffic count alone.
Do I need an expensive computer to run bee-counting computer vision at home?
Not necessarily. Lightweight detection models can run on affordable single-board computers like a Raspberry Pi, especially if you accept a lower frame rate or resolution. More demanding, higher-accuracy setups typically offload processing to the cloud instead.
How accurate are these systems at busy entrances during peak season?
Accuracy typically drops during very high-traffic periods because overlapping and occluding bees are genuinely harder to detect and track correctly. Most systems are tuned and validated more for reliable relative trend detection than for perfectly exact absolute counts during peak congestion.