A ghost target is any reported detection that does not correspond to an actual physical object or person in the scene. In people-counting or lab measurement contexts, ghosts cause inflated counts, erratic tracking, and misleading measurements.
Ghosts can appear as single false detections, duplicate tracks of one object, or transient noise spikes that look like a hit.
Radio or optical signals bounce off walls, glass, floors, or metal and arrive at the sensor with delays or altered angles, creating secondary returns that appear as separate objects.
Shiny or angled surfaces reflect energy nonlinearly. For depth cameras and LiDAR, specular reflections produce poor distance estimates that can be misidentified as objects.
Thermal noise, electronic noise, or weak signal strength increases false detections, especially at the edges of the sensor’s range.
Tracking and clustering algorithms may split one object into several tracks or interpret transient noise as a new target.
Moving HVAC vents, curtains, or hanging signs can intermittently trigger detectors if not accounted for in scene modeling.
Misaligned optics, incorrect timing, or poor calibration introduce systematic errors that look like ghosts.
Multiple sensors operating nearby (especially active sensors like LiDAR or radar) can create mutual interference and false returns.
Move sensors to minimize direct reflections from walls, glass, and polished floors. Avoid placing sensors facing large reflective surfaces or mirrors. Tilt sensors slightly downward or upward to change incident angles and reduce specularity.
Identify and, if possible, remove or mask moving fixtures that cause false triggers (e.g., hanging signs, curtains). Create a simple static background model so transient returns are easier to spot and dismiss.
Use adaptive thresholds that depend on signal strength and environmental noise level rather than fixed cutoffs. Require a minimum confidence or persistence (e.g., a detection must persist for N frames) before creating a new track.
Apply temporal filters (median, exponential smoothing) to reduce single-frame spikes. Use Kalman or particle filters for tracking so false transient returns don’t spawn persistent tracks.
Cluster returns spatially and reject detections that don’t match expected size or shape for the target class (e.g., human silhouette vs. point reflection).
Combine complementary sensors (e.g., thermal + radar, depth + thermal) so that inconsistent single-sensor returns are filtered out. Sensor fusion reduces dependence on one modality and helps disambiguate multipath artifacts.
If multiple active sensors are nearby, use time or frequency multiplexing to reduce mutual interference.
Improve pre-processing: denoise raw returns, remove outliers, and apply background subtraction. Use angle-of-arrival consistency checks for radar and phase-consistency checks for time-of-flight systems.
Run scheduled calibration routines to ensure timing, alignment, and gain settings remain accurate. Log raw sensor returns during tests to analyze ghost patterns and tune filters.
Use a checklist of scene layouts: empty room, single person at specific locations, multiple people moving predictable paths. Repeat tests at different times to capture environmental variability (lighting, HVAC cycles).
Introduce known reflectors (metal plates, glass panes) intentionally to observe ghost behavior and measure mitigation effectiveness.
Collect ground truth counts and positions manually or with a trusted reference sensor to measure false positive and false negative rates.
Track metrics: false positive rate, duplicate track rate, detection persistence, and latency. Use confusion matrices and time-series plots to identify when and where ghosts appear.
Apply one mitigation at a time and retest to isolate its effect. Keep a lab notebook of changes and outcomes.
Heat-based, low-resolution thermal sensing behaves differently from radar and depth systems in ways that can reduce certain ghost types while preserving privacy.
Practical takeaway: choose sensor modalities that match the environment and privacy requirements. For high-accuracy people counting in complex reflective scenes, thermal arrays combined with algorithmic filtering often yield fewer multipath ghosts than active optical systems.
Ghosts are most commonly caused by multipath reflections, sensor noise, and algorithmic splitting of a single object into multiple tracks.
Start with sensor placement and orientation, add temporal persistence checks, apply spatial clustering, and consider sensor fusion with a complementary modality.
Often software can mitigate many ghosts, but persistent issues from reflections or crosstalk may require hardware relocation, physical baffling, or different sensor types.
Yes, thermal sensors can produce false detections from reflected heat or environmental temperature changes, but they are generally less susceptible to optical multipath than optical depth sensors.
Use repeatable tests with ground truth, measure false positive and duplicate track rates, and iterate one mitigation at a time while logging raw sensor data.
Ghost detections arise from a mix of physics, hardware, and software issues. A structured approach — diagnose environment and hardware first, then tune software and add sensor fusion — produces the best results. For applications needing both accuracy and privacy, consider heat-based sensing as part of a fused solution to reduce common optical artifacts while preserving anonymous occupancy analytics.
West Coast, US
800 Airport Blvd
Burlingame, CA 94010
East Coast, US
501 Mass Ave
Cambridge, MA 02139
Japan
1 Chome-17-1 Toranomon, 16th Floor
Minato City, Tokyo 105-6415
