To solve the problems of high missed-detection rate and low recall rate existed in the YOLOv3 algorithm for detecting traffic lights, a traffic light detection method based on the optimized YOLOv3 algorithm is proposed. First, the K-means algorithm is used to cluster the data. By combining the clustering results with the statistical results of traffic light labels, the number and the width-height ratios of the prior boxes are determined. Then, the network structure is simplified according to the size characteristics of traffic lights. The 8× downsampling information and the 16× downsampling information are fused with high-level semantic information, and the object feature detection layer is established on two scales. Meanwhile, to avoid the disappearance problem of traffic light features with the deepening of the network, two sets of convolution layers are reduced before two object-detection layers, and thus the feature extraction steps are simplified. Finally, in the loss function, Gaussian distribution characteristics are used to evaluate the accuracy of the boundary box to improve the precision of traffic light detection. The experimental results reveal that the detection speed of the optimized YOLOv3 algorithm can reach 30 frames/s and the average precision is 9 percent higher than that of the original network, which effectively completes the detection of traffic lights.