YOLO Implementation and Performance Analysis Pipeline
- Raffay Hassan
- Feb 16
- 3 min read
Updated: Mar 30
Phase: 1 (Simulation Baseline)
Focus: YOLOv8n inference + lane-relevance filtering + performance reporting
Overview
This blog documents the YOLO implementation used , including the offline playback pipeline designed to evaluate detection behaviour and real-time feasibility. The aim was not only to run detections, but to quantify performance (latency/FPS) and reduce irrelevant detections using a lane-focused filtering approach.
Why YOLO in This Phase?
YOLO provides fast object detection and is commonly used in autonomous perception. The objective is to:
Run YOLOv8n across recorded scenario frames
Measure inference speed per frame
Track detection counts and class distribution
Reduce false relevance by isolating detections in the ego lane area
Generate evidence (CSV + graphs) suitable for dissertation reporting
Implementation Summary
The analysis tool loads:
Recorded RGB camera frames (camera/*.jpg)
Logged run outputs (sensor_fusion.csv, vehicle_state.csv)
LiDAR point clouds (lidar/*.ply)
It then performs:
Lane-area estimation from the image
YOLO inference on each frame
In-lane filtering of detections
Performance logging (ms/frame, FPS, detections/frame)
Output generation (CSV + graphs)
Lane-Relevance Filtering (Reducing Noise)
A key improvement is lane filtering: the system detects the drivable region and prioritises only objects that are likely relevant to collision risk.
Lane Mask Creation
Convert image to HSV
Apply colour thresholds targeting road-like surfaces
Apply trapezoid ROI (perspective lane region)
Use morphology (open/close) to clean the mask
In-Lane Object Test
For each bounding box:
take bottom-centre “ground contact” point
check pixel region around that point in the lane mask
accept detection if lane coverage exceeds a threshold
This enables two classes of outputs:
In-lane detections (highlighted, high priority)
Out-of-lane detections (faded, lower priority)
Performance Metrics Logged
For every frame, the pipeline logs:
Frame index
YOLO inference time (ms)
Number of detections (in-lane only)
Detected classes list per frame
Graph Outputs
The tool produces performance visualisations, including:
Inference time over frames
FPS over time
Objects detected per frame
Top detected object classes
Inference time distribution histogram
Detections vs inference time scatter
Cumulative detections over time
Summary box with mean/median/min/max stats
YOLO Inference Time
Mean inference time: ~9.2 ms
This corresponds to:
~114 FPS processing speed
One initial spike (~900ms) likely represents:
Model warm-up
First-time GPU memory allocation
After warm-up, inference stabilizes.
This shows:
YOLOv8n runs efficiently and consistently for real-time perception.
YOLO Processing Speed (FPS)
Mean FPS ≈ 114.5
This is significantly higher than:
CARLA simulation tick rate
Real vehicle control frequency
Meaning:
Object detection is not a bottleneck in the system.
Objects Detected Per Frame
Average ≈ 0.12 objects per frame
Why low?
Because:
I filtered detections to only objects in ego lane
Static obstacle was primary detection target
This confirms:
Lane-based filtering effectively removes irrelevant detections.
Top 10 Detected Classes
Dominant class: Car
Other minor detections:
Person
Stop sign
Random background objects
This shows:
YOLO correctly prioritizes relevant traffic participants.
Radar–LiDAR Agreement vs YOLO
The comparison graph shows:
Sensor fusion distance over time
YOLO detection counts per frame
Observation:
YOLO detects obstacle consistently when in camera view.
Fusion detects obstacle regardless of visibility.
During braking, both systems align temporally.
This demonstrates:
Camera perception complements sensor fusion but fusion ensures safety redundancy.
Outcome
This Phase 1 YOLO pipeline delivers:
Repeatable inference evaluation
Quantified performance evidence (ms/FPS)
Relevance filtering to reduce clutter
Report-ready CSV outputs and graphs
It acts as the perception evaluation foundation used to validate feasibility before expanding the project scope.
Comments