High visual recognition accuracy and accurate localization are the prerequisites for successful picking of grape picking robots, and high-precision localization of the picking point of grape bunches and clusters can effectively improve the picking success rate of the picking robot. However, grape bunches with multiple spike shapes and occlusion in natural scenes pose multiple challenges for recognition and localization. Aiming at the problem of low precision and inaccurate localization of grape bunches by picking robots in horizontal trellis-cultivated vineyard scenarios, we propose a method for detecting grape bunches and localizing picking points based on an improved YOLOv8 mode (YOLOv8-MH). Firstly, the detection ability of grapes with different spike shapes is strengthened by introducing the MSHA multi-head attention mechanism to improve the accuracy of the model, and some of the backbone networks are replaced with MobileNetV3 to reduce the number of model parameters and the amount of operations. Then, YOLOv8-MH is utilized to obtain the target detection results of grape bunches, and combined with the depth map-based region growing algorithm to obtain the critical points of grape bunches for picking point localization. The experimental results show that the average precision (AP) of the evaluation indexes of YOLOv8-MH reaches 93.1%, and the detection speed reaches 80 FPS, which is 2.7% higher than the precision before the improvement, and it has high detection precision for grapes with different spike shapes. Picking point localization tests were conducted on 30 groups of grape bunches in different scenarios, and the localization success rate was 94.4%, and the successfully localized picking points all met the localization tolerance of the end-effector. This study provides a necessary reference for the development of visual perception algorithms for grape picking robots.