Lightweight Deep Learning Model for Traffic Light Detection

Traffic Light Detection is one among the vital applications of Advanced Driving Assistant Systems (ADAS). The objective of traffic light detection is classification and localization of traffic lights present in the real-time images. State of art research outlines that the deep learning models have achieved better results than the traditional methods in detecting the traffic lights but still the techniques suffer from problems of low accuracy, slow speed, small object detection, illumination variations and occluded objects. In this paper Darknet19 feature extractor of You Only Look Once (YOLO) version 2 is replaced with SqueezeNet pretrained model. The. To improve the detection results, K-means clustering algorithm is applied for finding the clusters of bounding boxes of subset of LaRA traffic light dataset that comprises of green and red traffic lights and seven different sized anchor boxes were chosen for experimentation. Data augmentation techniques are also applied for increasing the diversity of the dataset. The proposed model attained mean average precision (mAP@.50) score of 84% and the visual results on the test dataset showed that the traffic lights were detected with increased confidence score with the proposed model in comparison with MobileNetV2, ResNet50 as backbones of the original YOLOv2 model.