In order to improve the accuracy of visual recognition and localization of used plastic labeled bottles, a lightweight binocular visual target detection algorithm based on the improved YOLOv5 is proposed. Firstly, in the backbone network, the original CBS and C3 modules are respectively replaced by the GhostBottleNeck module and the GhostNetV2 module. Secondly, GSConv and GSConv-based VOVGCSP modules are introduced into the neck network. The dataset is trained by field real measurements to be used for training the improved YOLOv5 model. Based on the optimized algorithm, a binocular camera is used to investigate the used plastic bottle ranging system. Results show that 1) the improved YOLOv5 model increases the accuracy from 87.92% to 95.39%, the number of parameters decreases from 7 012 888 to 5 933 320, and the FPS improves from 96 frame/s to 105 frame/s when comparing with the original model. The replacement network effectively reduces the number of parameters of the model, and the accuracy of the improved model improves by 7.47%. 2) The maximum error of detection is about 7 mm through binocular ranging experiment, and the relative error is within 1%, which meets the requirements. The identifying and locating accuracy both meet the requirements of real-time processing, which can provide technical support for the development of intelligent recycling equipment.