Abstract
Estimating the heights of objects in the field of view has applications in many tasks such as robotics, autonomous platforms, and video surveillance. Object height is a concrete and indispensable characteristic that machines must capture to perform actions like obstacle avoidance. Traditionally, object height is estimated using laser ranging, radar, or stereo cameras. However, the cost of these sensors can inhibit their adoption. Our approach requires only a single 2D image. We introduce the Monocular Object Height Estimation Network (MOHE-Net), which includes a cascade of two networks. The first, OD-Net, performs object detection to identify bounding boxes. This information is then input to HE-Net, a linear Multi-layer Perceptron (MLP) that models camera-scene geometry to estimate height. Unlike standard MLPs, HE-Net does not require training or activation functions. The approach works for both static and moving platforms, achieving state-of-the-art performance suitable for real-time autonomous navigation.
Introduction: The Scale Problem in Monocular Vision
Determining the physical height of objects from a single camera—without the help of expensive LiDAR or stereo setups—is a classic "hard" problem in computer vision. Because a 2D image loses depth information, a small object close to the lens can look identical to a large object far away. This ambiguity, known as the scale problem, makes 2D-to-3D projection mathematically underspecified without additional constraints.
Existing monocular methods often rely on complex depth-map estimation or continuous camera pose tracking, which can be computationally expensive and prone to drift. MOHE-Net aims for a middle ground: high accuracy and real-time speed by leveraging known geometric constraints of the environment.
The MOHE-Net Architecture
The system uses a two-stage pipeline to bridge the gap between 2D pixels and 3D metric measurements:
- OD-Net (Object Detection): A standard pre-trained convolutional neural network (such as YOLOv5 or Faster R-CNN) identifies objects and draws bounding boxes around them.
- HE-Net (Height Estimation): This is the core innovation. Instead of a traditional "black box" neural network that learns from massive amounts of data, HE-Net is a handcrafted linear Multi-layer Perceptron (MLP).
Scene Geometry and the Linear MLP
The "weights" of the HE-Net MLP are not learned through backpropagation; they are derived directly from the camera's intrinsic parameters and the scene's geometry. The model relies on two key physical constraints: the camera views a piece-wise flat ground (Planar Scene), and objects stand vertically on that ground (Upright Objects).
By using a one-time calibration with ground markers (Direct Linear Transformation or DLT), the network estimates the depth ($z$) of the object's base. It then back-projects the 2D bounding box points into 3D camera coordinates. The physical height is calculated as the vertical distance between the 3D top and bottom points.
In deep learning, linearity is usually a limitation, but here it is a feature—the linear MLP perfectly mirrors the linear nature of projective geometry. Because the weights are determined by calibration, HE-Net requires no training data for the height estimation component itself.
Experimental Evaluation
The researchers evaluated MOHE-Net using a custom dataset collected from a monocular camera mounted on a moving vehicle across three scenarios: stationary objects, a walking person (to test gait effects), and a moving platform in a parking lot. Calibration was established using a grid of 36 markers to define the homography matrix.
Key Results
- Accuracy: In tests involving a walking person, the system achieved a mean error of just 5.08 cm (approximately 2.78% relative error).
- Speed: When paired with YOLOv5x, the system ran at 26.4 FPS on an NVIDIA Titan V, making it suitable for real-time deployment.
- Versatility: The system successfully estimated heights for a wide variety of objects, ranging from small sports balls (25 cm) to large vans (233 cm).
Conclusion
MOHE-Net demonstrates that combining modern object detection with rigorous geometric modeling can solve complex 3D tasks without the need for expensive sensors. While the system is currently limited by the "flat world" assumption and sensitivity to occlusions, it provides a robust, high-speed framework for obstacle avoidance in autonomous platforms. Future work will focus on adapting the model to uneven terrain and improving robustness against low-contrast backgrounds.
Summary updated