Performance Evaluation of Image Enhancement Techniques for Traffic Cone Detection Under Twilight Conditions Using YOLOv8

Abstract

This paper presents the performance evaluation of various image enhancement techniques applied to traffic cone detection using YOLOv8 architecture. A real-world simulation was conducted by placing traffic cones along both sides of a roadway to serve as visual warning signals in twilight conditions. Data collection was performed through real video recording using a smartphone camera mounted on the car’s dashboard. The camera features a CMOS sensor and a wide-angle lens, capturing images at a resolution of 1920x1080 pixels during twilight hours (18.30–19.00). Ambient illumination levels typically range between 1 and 10 lux, resulting in significantly reduced visibility and image clarity. A total of 900 video frames were extracted and organized into one base dataset (SRC) and then duplicated into five additional sets. They were processed using five enhancement techniques: 1) Linear Adjustment (LIN), 2) Histogram Equalization (HE), 3) Logarithmic Enhancement (LOG), 4) Linear Adjustment with Saturation Handling (LIN-SAT), and 5) Contrast Limited Adaptive Histogram Equalization (CLAHE). Traffic cone detection was performed using a YOLOv8 model specifically trained for this task. Evaluation metrics included Average Cone Count per frame, Precision, Recall, F1 Score, Accuracy, and Average Confidence. Experimental results showed that CLAHE yielded the best overall performance, detecting an average of 4.7 cones per frame with 98.74% Precision, 77.78% Recall, an F1 Score of 87.01%, 77.01% Accuracy, and 62.22% Average Confidence. The results confirm that CLAHE significantly improves the robustness and accuracy of YOLOv8-based traffic cone detection under low-light conditions.