Dynamic-TLD: A traffic light detector based on dynamic strategies

Traffic light detection technology can assist drivers in making decisions and has potential applications in autonomous driving to reduce loss of life and property. However, the detection of traffic light objects is challenging due to traffic light’s small size and complex backgrounds, making it difficult for current technology to meet real-time and high accuracy requirements. In this work, we identify that inappropriate feature integration and static parameters during detector training limit model efficiency. Specifically, using the same detection criteria for all object sizes in the feature integration process reduces the effectiveness on small objects such as traffic light. Additionally, a static IoU threshold during training results in insensitivity to objects in complex backgrounds. Finally, static parameters in the loss function hinder the learning of high-quality samples. To address these problems, we propose Dynamic-TLD. This approach improves the training efficiency of small samples through the weighted integration of low-level and high-level features. It also enhances the learning of traffic light images by gradually increasing the IoU threshold based on model learning. Additionally, the learning of high-quality samples is facilitated by dynamic parameter selection in the loss function. Compared with six other advanced detectors, our approach improves average precision on three publicly available traffic light datasets by 1.4 points, 2.1 points, and 0.8 points respectively while meeting the real-time requirement of autonomous driving systems. These results demonstrate the effectiveness and robustness of our approach and its potential for application in autonomous driving systems. Codes and demonstration video are available at https://github.com/ljyw17/traffic-light-detector.