A Deep Learning Method for Traffic Light Status Recognition

Abstract

Real-time and accurate traffic light status recognition can provide reliable data support for autonomous vehicle decision-making and control systems. To address potential problems such as the minor component of traffic lights in the perceptual domain of visual sensors and the complexity of recognition scenarios, we propose an end-to-end traffic light status recognition method, ResNeSt50-CBAM-DINO (RC-DINO). First, we performed data cleaning on the Tsinghua-Tencent traffic lights (TTTL) and fused it with the Shanghai Jiao Tong University's traffic light dataset (S2TLD) to form a Chinese urban traffic light dataset (CUTLD). Second, we combined residual network with split-attention module-50 (ResNeSt50) and the convolutional block attention module (CBAM) to extract more significant traffic light features. Finally, the proposed RC-DINO and mainstream recognition algorithms were trained and analyzed using CUTLD. The experimental results show that, compared to the original DINO, RC-DINO improved the average precision (AP), AP at intersection over union (IOU) = 0.5 (AP 50 ), AP for small objects (AP s ), average recall (AR), and balanced F score (F1-Score) by 3.1%, 1.6%, 3.4%, 0.9%, and 0.9%, respectively, and had a certain capability to recognize the partially covered traffic light status. The above results indicate that the proposed RC-DINO improved recognition performance and robustness, making it more suitable for traffic light status recognition tasks.