State of the Art - Visual Odometry

Visual odometry is a technique that estimate robot location using visual information (images), the position is defined as the incremental estimation of robot motion from image sequences using an on-board camera. Visual odometry is a method highly used on off-road robots like space robotic exploration missions. The motion of the robot is estimated with calculation of the pixel displacement between two consecutive frames, this is called optical flow. [1]

The optical flow is mostly calculated based on features that stand out in the image, this method doesn't work well on smooth or texture-less surfaces like concrete and asphalt roads. The other method, with better results, is template matching. This method use a small area of a image and tries to find it in the next frame, if found it is possible to calculate the optical flow. [2]

There are two main different approaches in hardware, the first uses a camera looking forward and the velocity is calculated based on the movement of the surroundings. This method is very sensitive to lighting changes and a high quality image is needed. The other method uses a camera looking directly at the ground, this is the principle used in most optical mice. This method needs simpler computation and hardware and lighting conditions can be handled better. [3]

The literature suggests visual odometry as a better odometry system in terms of precision and error accumulation. Navid Nourani-Vatani et al [2], [4] obtained good results using a common webcam at 20 fps and a image resolution of 640 x 480 pixels, they used template matching with a searching area of 320 x 320 pixels (figure 1). The acquisition speed was set to 20 fps due to the high processing time, 42 ms/frame, this is a clear limitation. Also due to the low frame rate, the velocity of the vehicle was limited to 1.5 m/s, which is very low. Nevertheless this method showed better results than common wheel's odometry.

The velocity allowed by the measuring unit is are mostly influenced by the sampling rate and the size of the image. If the velocity is higher than the limit, there will be no overlap of the frames and there will be no similarity between the consequent images, this will cause the impossibility of the estimation of velocity.[3] So for car-like velocities it is necessary a very high sampling rate camera, this brings difficulties as the computational power needs to be considerable. Viktor Kálmán [3] used line scan cameras to achieve high velocities and resolution, using less computational resources than a common matrix camera. The sensor used was capable of measuring movements of vehicles up to the range of 100m/s with high accuracy.
Visual odometry has clear advantages over the common wheel's odometry. It is possible to measure velocity independent of the variations in tire pressure, tire diameter, uneven terrain and tire slip. It is possible to obtain a compact and easy to install module. However there are several drawbacks, the lighting is an important factor and it isn't always possible to control, also the sensors are high price and sophisticated.

[1] - "Visual Odometry on the Mars Exploration Rovers" by YANG CHENG, MARK W. MAIMONE, AND LARRY MATTHIES

[2] - "IMU Aided 3D Visual Odometry for Car-Like Vehicles" by Navid Nourani-Vatani, Jonathan Roberts and Mandiam V. Srinivasan

[3] - "On modeling and control of omnidirectional wheels" by Viktor Kálmán

[4] - "Practical Visual Odometry for Car-like Vehicles" by Navid Nourani-Vatani, Jonathan Roberts and Mandiam V. Srinivasan