Accurate estimation of joint angles in industrial manipulator robots can help to improve robotic applications in different scenarios. Depending on the robot's size and application, as well as the actuator type (it can be pneumatic) it can be expensive and hard to acquire an encoder to obtain feasible joint angle information. Moreover, visual joint angle estimation for manipulators can be used for robot automatic calibration, anomaly detection, and monitoring. In this work, we propose a joint angle estimation for an ABB industrial manipulator robot by using convolutional object detection methods (CNNs) and k-means clustering. In particular, we trained and evaluated the use of the You Only Look Once (YOLO) algorithm for the joint angle detection task. For this, we first build an image dataset of different joint configurations of the manipulator robot. Then, we train, calibrate, and test the YOLO object detection method. Then, we analyze several frames to obtain different detection measures and apply k-means clustering to improve the joint position estimation. Finally, we employ geometric analysis to estimate the joint angles considering a lateral perspective of the robot. We compare the estimation of the joint angles with the encoder measures of the robot, and we demonstrate that our method can reach low error results at different robot configurations. The proposed method can enhance joint angle estimation for different industrial applications.