Object detection is crucial for autonomous robotic systems to interact with the world around them but, in robots with low computational resources, deep learning is difficult to take advantage of. We develop incremental improvements to related work on feature approximation and describe an adaptive fHOG feature pyramid construction scheme based on histogram downsampling, together with a SVM classifier. Varying the pyramid level to which the scheme is applied gives control over the trade-off between precision or speed. We evaluate the proposed scheme on a modern computer and on a NAO humanoid robot in the context of the RoboCup competition, i.e., robot and soccer ball detection, in which we obtain significant increase (1.57x and 1.68x on PC and robot respectively) in pyramid construction speed relative to our baseline (the dlib library) without any loss in detection performance. The scheme can be adapted to increase speed while trading off precision until it reaches the conditions of a state-of-the-art power law feature scaling method.