Key Performance Indicators
Simulated evaluation dataset for single-device prototype performance analysis
Total Detections
5,000
Detection Accuracy
92.56%
Avg AI Confidence
0.926
Avg Response Time
145.9ms
High-Risk Alerts
1,373
Avg Battery
73.9%
Prototype Device
1
Night Evaluations
1,483
Daily Detection Volume & Accuracy Trend
Simulated detection records per day with accuracy overlay · Prototype Evaluation Dataset
Monthly Detection Trend & Accuracy
Volume by month with detection accuracy %
Detection Status Breakdown
Successful · Missed · Repeated across all sessions
Distance Categories
How close objects are when detected
Alert Level Distribution
High · Medium · Low risk events
Object Position in Frame
Left · Center · Right classification
Key Insight: Detection accuracy trends from 91.6% (Dec 2025) to 94.4% (Jun 2026) across the simulated evaluation dataset, indicating strong and stable model performance. Person is the dominant detected class (35.6% of all records). 30% of detections are flagged as Very Close — validating the prototype's proximity alerting logic on a single Raspberry Pi 4 running MobileNet-SSD.
Top 10 Detected Object Classes
Detection count with average AI confidence per class
Detection Confidence Distribution
Histogram of confidence scores across 5,000 detections
Detections by Hour of Day
Activity heatmap with accuracy overlay (24-hr cycle)
High-Alert Events by Object Class
Which objects trigger the most urgent alerts
Object Detection – Detailed Summary
Class-level count, confidence, response time, and risk profile
Detection Volume by Location Type
Where detections occur most frequently
Day vs Night Performance
Environment impact on AI confidence
Key Insight: Person accounts for 35.6% of all detections and carries the highest AI confidence (0.944). Plant and Car show the lowest confidence (0.883 / 0.888), suggesting these classes benefit most from future retraining. Night-condition records maintain near-identical accuracy to day-condition records in the simulated dataset, demonstrating the prototype's expected low-light robustness.
High Alerts
1,373
Medium Alerts
1,635
Low Alerts
1,992
Very Close Events
1,500
Monthly Alert Breakdown
High / Medium / Low alert counts per month
Alerts by Location Type
Risk distribution across environment categories
Session Duration Distribution
Minutes per session (all 5,000 events)
Detection Accuracy by Location
Where is the system most reliable?
Daily High Alert Events Over Time
Frequency of High severity triggers per day
Key Insight: Outdoor and Indoor scenario groups produce the highest volume of high-priority alerts in the simulated dataset, while School scenarios yield the best accuracy (94.4%). Short-duration records (0–10 min) dominate the dataset, reflecting brief navigation-style evaluation scenarios.
Detections per Scenario Group
Simulated detection records grouped by evaluation scenario (single prototype device)
Detection Accuracy per Scenario Group
Accuracy % per scenario group — reflects prototype performance across varied conditions
Average Battery Level per Scenario Group
Battery % recorded across evaluation scenarios · prototype device
Average Response Time per Scenario Group
Milliseconds from detection to voice alert across evaluation scenario groups
Prototype Evaluation — Scenario Performance Matrix
20 evaluation scenario groups · Single prototype device (Raspberry Pi 4 + MobileNet-SSD) · Simulated dataset
Key Insight: This simulated evaluation dataset models 20 scenario groups on a single prototype device to analyse performance across varied object types, environments, and distances. Scenario S-07 and S-13 recorded the highest detection counts (261 & 276 records) while S-07 also achieved the best accuracy (95.4%). S-01 and S-17 show slightly lower accuracy (90.2–90.3%), reflecting performance variation across different evaluation conditions. Battery remained healthy across all scenarios (71–76% average).
Avg AI Confidence
0.926
Best Class Confidence
0.944
Avg Response Time
145.9ms
Detections > 0.95
55.1%
AI Confidence Score Distribution
Full histogram across all 5,000 detection events
Confidence by Object Class
Average model confidence for each detected category
Monthly Average Confidence Trend
Has model confidence improved over deployment?
Response Time by Object Class
Which classes trigger voice alerts fastest?
Daily Average Confidence — Prototype Evaluation Dataset
MobileNet-SSD confidence stability across all 5,000 simulated detection records (single prototype device)
Key Insight: The MobileNet-SSD model maintains remarkably stable confidence (0.903–0.945 range daily) across 182 days, with 55.1% of detections scoring above 0.95. Plant (0.883) and Car (0.888) are the lowest-confidence classes — prioritising these in future training data collection will yield the most accuracy gains.