AI Traffic Counting vs Manual Counting: An Honest Comparison

Manual traffic counting has been the default method for traffic surveys in South Africa for decades. AI-powered video analysis is increasingly competitive — but it is worth looking honestly at where AI wins, where it struggles, and where the two methods should be combined.

Cost comparison

A standard 12-hour manual TMC survey at a four-way intersection typically costs between R2,000 and R5,000 in South Africa when factoring in field observer time, supervision, office processing, and report preparation. That cost scales linearly: 10 intersections means 10 × the cost.

Video-based AI counting has a different cost structure. The camera setup cost is a one-time expense. Processing cost per video is low — TallyRoad charges by processing hours rather than per survey. The result is that AI becomes significantly cheaper at scale, especially when surveys need to be repeated (before-and-after studies, seasonal variation analysis).

Accuracy

This is where honest discussion matters. AI detection accuracy depends heavily on video quality and conditions.

For motorised vehicles (cars, trucks, buses, motorcycles) in clear daytime conditions with a fixed overhead camera, modern AI models including YOLOv11 achieve accuracy above 95% for detection and counting. This is comparable to careful manual counting.

Accuracy drops in difficult conditions: night-time footage without infrared, heavy rain, fog, severe occlusion (vehicles hidden behind each other), and oblique camera angles. Manual counting handles these better because humans can make inferences that current AI models cannot.

Pedestrian and cyclist counting is also harder for AI — smaller objects, more varied behaviour, higher occlusion rates.

Turnaround time

Manual counting: field survey happens in real time, but data processing and report preparation typically takes another day or two. Total time from survey to report: 2–5 days.

AI counting: once the video is uploaded and configured, processing takes minutes to an hour depending on video length and model size. A 2-hour intersection video is typically processed in under 15 minutes on GPU.

Flexibility

Manual counting has a significant limitation: if you need to change the counting methodology (different time intervals, different vehicle classes, different approach directions), you need to redo the field survey.

AI counting lets you reprocess the same video with different settings. Forgot to draw a particular approach line? Reprocess. Client wants 5-minute intervals instead of 15? Reprocess. This reprocessability is a genuine advantage that changes how traffic engineers work.

When to use each

Use AI video counting when: you have existing video infrastructure, you need rapid turnaround, you are doing before-and-after studies, or you are processing multiple sites simultaneously.

Use manual counting when: you do not have video coverage, lighting or weather conditions are too poor for camera-based counting, or you need to count non-standard movements that AI does not yet classify.

The optimal workflow for many engineers is AI-first: set up a camera, let AI handle the baseline counts, and only deploy manual observers for edge cases or quality verification.