Sources and optimization methods of behavioral analysis false positives

Six major sources of false positives

1. Light and shadow changes: tree shadows swaying, clouds blocking, water reflections, car lights sweeping across. This is the most common source, especially in perimeter scenes.
2. Small animals: Cats, dogs, and birds trigger boundary crossing or area invasion.
3. Foliage and floating objects: Wind-blown branches, plastic bags, flags.
4. Weather: Rain, snow, and haze will change the characteristics of the picture, and the impact will be greater at night.
5. The target is too small or blocked: There are insufficient pixels for distant targets, or the crowd blocks each other, leading to misjudgment.
6. Improper setting of the rule area: The area is too large, the boundary is close to the edge of the screen, or it contains interference sources.

Optimization steps

1. Delineate effective detection areas: Exclude interference areas such as trees, water, roads, etc., and only retain the areas that really need to be monitored.
2. Set target size filtering: Ignore targets that are too small (such as moving objects smaller than 1% of the frame), and can filter out most small animals.
3. Set time policy: Use different sensitivities during the day and night; temporarily block during construction periods and peak hours.
4. Adjust sensitivity grading: First set a higher threshold to reduce false positives, and then gradually lower it to find false negatives, which is more efficient than the other way around.
5. Enable target type filtering: Only alert "people" or "cars", which can greatly reduce interference from animals and floating objects.
6. On-site iterative tuning: Analyze one by one based on real alarms for 2 to 4 consecutive weeks and continue to optimize.

Suggestions on acceptance indicators

Instead of requiring "zero false positives", it is better to set measurable indicators:

• Perimeter scenario: False alarms ≤ 2 times/(week·defense zone), false alarms ≤ 1 time/month.
• Indoor scene: false alarm ≤ 1 time/(day·channel).
• Observe continuously for 7 days, and calculate the average daily alarm volume and effectiveness (proportion of real events).
• The recommended target for effectiveness is ≥ 80%, that is, at least 8 out of 10 alarms are real events.

Supporting means to reduce false positives

• The perimeter adopts "infrared beam/grating + video review" dual detection, and the false alarm rate can be an order of magnitude lower than that of pure video.
• The alarm is linked to the voice announcement, which creates a deterrent and facilitates rapid manual judgment.
• Establish an alarm review process to prevent on-duty personnel from ignoring alarms due to too many false alarms.