Measuring Traffic
Jana Mohamed, Charlotte Kyer, Addison Massey
2026-10-13
Analysis of Residential Traffic Patterns Using a Raspberry Pi
Over just a two-week period, a subsection of Drew Hill Ln in Chapel Hill, North Carolina was observed with over 74,000 object counts, including cyclists, pedestrians, cars, heavy vehicles, and unidentified. Drew Hill Ln is a small public road that primarily serves Notting Hill apartment residents and visitors. These observations were recorded continuously from September 23, 2026 to October 10, 2026 on a Raspberry Pi. On October 6th and 7th, construction occurred on the recorded section of the street, significantly disrupting recorded traffic patterns during that period. This report analyzes how traffic patterns changed during and following this period, and how the methodology impacts and limits the validity and accuracy of these findings.
The map above shows the approximate location of the sensor for the duration of the data collection period.
Overview of the Data
The initial analytical portion of this narrative, focused on reviewing records with class labels which narrowed our object count down to 18,696. The below tables depict a summary of these records broken down by class type, number of observations, share of the data, and average confidence of the classification.
| Measure | Count |
|---|---|
| total_records | 74156 |
| missing_labels | 55460 |
| missing_confidence | 55460 |
| missing_label_or_confidence | 55460 |
| usable_records | 18696 |
| Road-user type | Number of records | Share of records (%) | Average confidence (%) |
|---|---|---|---|
| Car | 11756 | 62.9 | 74.6 |
| Pedestrian | 5163 | 27.6 | 67.9 |
| Heavy vehicle | 1662 | 8.9 | 77.4 |
| Cyclist | 115 | 0.6 | 61.5 |
Comparing Classification Confidence Across Road-User Types
The average confidence of identified by typer anges from 61.5% to 77.4%, a range that could suggest limitations of our data and its collection process. To evaluate this idea, we started by comparing the average classification confidence for each type of road user and the number of records for each category. The aim was to provide theories as to the cause of variations in road-user types confidence and whether there was a relationship between confidence and the number of records.
These figures showed that cyclists had the lowest average confidence at 61.5%, along with the fewest records, though there did not appear to be a strong correlation to observation counts. This stood out in the data because a residential area would expect to see elevated cyclist activity.
Investigating Possible Missed Detections
Our next question was: Are cyclists being missed entirely, or are they being detected but classified as pedestrians or another road-user type? To research this, cyclists records were grouped by date then summarized.
| Date | Records labeled cyclist |
|---|---|
| Sep 02, 2026 | 0 |
| Sep 21, 2026 | 0 |
| Sep 23, 2026 | 1 |
| Sep 24, 2026 | 6 |
| Sep 25, 2026 | 3 |
| Sep 26, 2026 | 1 |
| Sep 27, 2026 | 1 |
| Sep 28, 2026 | 2 |
| Sep 29, 2026 | 1 |
| Sep 30, 2026 | 6 |
| Oct 01, 2026 | 3 |
| Oct 02, 2026 | 3 |
| Oct 03, 2026 | 0 |
| Oct 04, 2026 | 0 |
| Oct 05, 2026 | 2 |
| Oct 06, 2026 | 51 |
| Oct 07, 2026 | 32 |
| Oct 08, 2026 | 0 |
| Oct 09, 2026 | 3 |
| average_per_day | median_per_day | highest_daily_count |
|---|---|---|
| 6.05 | 2 | 51 |
These tables show large variation in counts per day. To better view the distribution of cyclists over the study period, a graph was created of daily records with the number of recorded cyclists.
Checking Cyclist Images and Missing Classification Labels
Given this variation across the observation period, we reviewed images in the Google Drive folders, especially the No Detection folder, to look for examples that could help explain the results. Ultimately, no cyclists were found in the No Detection folder, so focus was then diverted to the cyclist category, where manual verification of the images from several dates occurred. Multiple images were found containing visible cyclists and then matched with their corresponding records in the dataset. Some of these records had NA for class_label. This means the Pi had saved a record associated with a visible cyclist, but the record did not have a classification label. This raised another possibility: the number of cyclists recorded in the dataset may be lower than the actual number because some cyclist-related records have missing labels. To investigate this further, we manually counted visible cyclists on three dates and compared our counts with the number of records labeled as cyclists in the dataset.
| Date | Manual cyclist count | Pi cyclist records | Difference |
|---|---|---|---|
| Sep 25, 2026 | 2 | 3 | 1 |
| Sep 26, 2026 | 2 | 1 | -1 |
| Oct 05, 2026 | 4 | 2 | -2 |
| Oct 06, 2026 | 6 | 51 | 45 |
The peak on October 6, 2026 reached 51 cyclist-labeled records, compared to only 6 cyclists counted manually. Reviewing the images revealed that some records were actually construction vehicles, suggesting misclassification contributed to the spike.
| id | time_start | class_label | confidence | direction |
|---|---|---|---|---|
| 53802 | 2026-10-06 8:53:57 | cyclist | 0.5889058 | RL |
| 58654 | 2026-10-06 11:58:04 | cyclist | 0.5129679 | LR |
The camera classified a Caterpillar skid steer loader and a Caterpillar mini excavator as cyclists. These examples suggest that the model may struggle to distinguish unusual construction vehicles from cyclists.
Visual Inspection of Classification and Detection Issues
While sorting through the images, we noticed another concern: repeated captures of the same vehicle. The Raspberry Pi sometimes captures the same vehicle in multiple images as it moves through the camera’s view. These repeated captures can increase the number of records without representing additional vehicles. Below depicts a slow-moving car that was counted twice in the records within a span of three seconds.
| id | time_start | class_label | confidence | direction |
|---|---|---|---|---|
| 73862 | 2026-10-09 13:23:04 | heavy_vehicle | 0.8871141 | RL |
| 73863 | 2026-10-09 13:23:07 | heavy_vehicle | 0.9080949 | LR |
Classification of Unusual Vehicles
The classification also sometimes inappropriately identified objects. A serious of images below show objects incorrectly labeled in their matching record.
| id | time_start | class_label | confidence | direction |
|---|---|---|---|---|
| 1538 | 2026-09-23 15:41:47 | heavy_vehicle | 0.7772031 | LR |
| 52174 | 2026-10-05 11:20:49 | heavy_vehicle | 0.8022027 | LR |
| 73843 | 2026-10-09 13:08:51 | heavy_vehicle | 0.6995968 | RL |
While all three golf carts were classified as heavy vehicles, this is likely not the most appropriate category for golf carts. This may be a limitation of the camera’s classification system, since its categories are broad and may not include unusual types of road users. As a result, vehicles that don’t fit neatly into the available categories may be misclassified.
| id | time_start | class_label | confidence | direction |
|---|---|---|---|---|
| 52527 | 2026-10-05 15:12:51 | car | 0.9021476 | RL |
The camera classified a scene with a person standing beside a car as a car, with 90.2% confidence.
| id | time_start | class_label | confidence | direction |
|---|---|---|---|---|
| 52665 | 2026-10-05 18:13:36 | car | 0.7960538 | LR |
The camera classified a tow truck carrying a car as a car, with 79.6% confidence.
| id | time_start | class_label | confidence | direction |
|---|---|---|---|---|
| 1469 | 2026-09-23 15:17:29 | car | 0.8764458 | RL |
The camera correctly classified the image of two cars as a car, with 87.6% confidence.
| id | time_start | class_label | confidence | direction |
|---|---|---|---|---|
| 11329 | 2026-09-25 12:51:43 | pedestrian | 0.66724 | RL |
The camera classified an image of four people as a pedestrian, with 66.7% confidence.
Analyze Construction Vs. Non-Construction
During our data collection period, construction occurred within our camera’s view on October 6th through 7th, leading to unexpected results in our counts of traffic along the road. Initially, we were concerned that this disruption would pose challenges for our project; however, we realized that it gave us a unique example of how construction disrupts normal travel behaviors within residential areas.
Pedestrian and heavy vehicle counts increased during the construction period, and pedestrian counts in particular drastically increased. To visualize this event, we first identified observations within the construction time period to compare their total and average daily traffic counts across classification type.
We first compared total average traffic counts between construction and non-construction days, and found that there was a fairly large increase in pedestrian and heavy vehicle observations on construction days. We anticipated that both of these types would be recorded more often on construction days due to the presence of construction vehicles and workers walking back and forth. Naturally, heavy vehicle counts were higher during construction times. We were concerned that car counts would be artificially higher than average on construction days than on non-construction days due to work trucks being parked and moved around. While these artificial car observations are likely to factor in for some of our counts, our data found that, on average, fewer cars were counted on construction days than on non-construction days.
| construction | class_label | total_count |
|---|---|---|
| Construction | car | 1089 |
| Construction | cyclist | 83 |
| Construction | heavy_vehicle | 1180 |
| Construction | pedestrian | 4194 |
| Construction | NA | 9213 |
| Non-Construction | car | 10667 |
| Non-Construction | cyclist | 32 |
| Non-Construction | heavy_vehicle | 482 |
| Non-Construction | pedestrian | 969 |
| Non-Construction | NA | 46247 |
| construction | class_label | mean_daily_count |
|---|---|---|
| Construction | car | 544.500000 |
| Construction | cyclist | 41.500000 |
| Construction | heavy_vehicle | 590.000000 |
| Construction | pedestrian | 2097.000000 |
| Construction | NA | 4606.500000 |
| Non-Construction | car | 711.133333 |
| Non-Construction | cyclist | 2.666667 |
| Non-Construction | heavy_vehicle | 32.133333 |
| Non-Construction | pedestrian | 57.000000 |
| Non-Construction | NA | 2720.411765 |
To better understand this change, we calculated the percent change of pedestrian and vehicle types during constructions and non-construction periods. There was a 23% decrease in car activity, but a 1456% increase in cyclists, 1736% increase in heavy vehicles, and 3579% increase in pedestrians. These alarmingly high numbers indicate that we are most likely correct in assuming construction activity artificially inflated our collected data.
| class_label | Construction | Non-Construction | pct_change |
|---|---|---|---|
| car | 544.5 | 711.133333 | -23.43208 |
| cyclist | 41.5 | 2.666667 | 1456.25000 |
| heavy_vehicle | 590.0 | 32.133333 | 1736.09959 |
| pedestrian | 2097.0 | 57.000000 | 3578.94737 |
| NA | 4606.5 | 2720.411765 | 69.33098 |
Despite these challenges, we still wanted to understand just how construction impacted people’s usage of this residential road. While we knew that the counts during the construction period were likely unreliable, we wanted to know how travel behaviors differed before and after the construction. We split the dates into time periods and calculated the averages per classification type. This found that car travel decreased after construction, meaning that people were aware that construction was happening in that corridor and wanted to avoid it if they could. However, we found that pedestrian, cyclist, and heavy vehicles increased after construction. An idea for this observation based on images we looked at was that office staff and construction workers came by for the next couple of days to check on the work and remove cones. Similar to the previous pedestrian issue, the sensor may have picked up on these people walking into and out of the frame and counted it as multiple observations. However, another idea for the increase is that people may have been curious about what construction was happening and decided to use their morning walk to snoop.
Further Limitations and Conclusion
Raspberry Pi data collection can raise several accuracy issues. The environmental conditions associated with the collection process, including glare, lightning, weather, setup, and position, can all impact the quality of the records. In this report, both under- and over-classification of labels occurred on various days, reflected in object confidence. Abnormal objects or activities can also confuse the Pi’s algorithm. In our case, construction days proved to have strong outliers in counts resulting from misclassification of construction vehicles. However, these results do not have to be fully discounted; we can still draw a general trend from this report. Despite overclassification of some classes, construction likely affected the typical travel behavior of individuals within this study area.