Obstacle detection is a key capability for autonomous vehicles, and especially important one for vehicles navigating in agricultural environments. In essence, obstacle avoidance consists of determining whether the space ahead of the vehicle is clear from obstructions for safe travel. Its goal is to detect all obstacles along the path in time for the vehicle to react to them, while minimizing misclassifications.
An automated tractor that operates on a farm needs to detect obstacles in order to; provide safety for humans that get too close to the vehicle; avoid causing damage to the environment (by collisions with trees, tools or other equipment located on the vehicle’s path); and avoid damaging or incapacitating itself due to ditches, irrigation canals, rocks, gopher holes, etc. These three factors make false negatives expensive; as a result, having a reliable and robust obstacle detection system is a hard prerequisite for any kind of automation in a real world setting.
Unfortunately, the fact that the natural environment contains very little structure and presents a large number of uncontrollable factors makes outdoor obstacle detection very difficult. The difficulty is proportional to the generality required from the system. While finding obstacles on very flat terrain is easily solved, creating a general system that will work in row crops, orange groves and other agricultural settings at any time of the day and in any weather can prove to be extremely complex.
Part of the problem is due to the fact that no sensor exists that can guarantee detection in the general case. Each sensor has failure modes that can make the automated vehicle unsafe unless there are other sensors that can cover these failure modes. For camera-based sensing, the changes in natural light that occur during the day, the limited dynamic range of the cameras and the various algorithmic limitations (like the lack of texture for stereo vision) significantly diminish obstacle detection capabilities. Another sensor commonly used for obstacle detection is a laser range finder, with either one- or two-axis scanning motion. But two-axis lasers can be very expensive and slow, if we take into consideration the requirements imposed by vehicle motion; one-axis lasers are cheaper and faster, but this advantage comes at the expense of significantly less data than provided by a two-axis laser.
Obstacle detection systems have been developed for cross country navigation. The most common technique employed uses neural networks for terrain classification, operating on raw RGB (Red, Blue, Green) data or on features, such as intensity variance, directional texture, height in the image, and statistics about each colour band. The MAMMOTH system drove an off-road vehicle. MAMMOTH used two neural networks (with one hidden layer each) to analyze separately visual and laser rangefinder data. Each neural network output a steering command. A “task neural network” took as inputs the outputs of the hidden layers in the two networks and fused the information into a steering command. The JPL system detected obstacles in an off-road environment by combining geometric information obtained by stereo triangulation with terrain classification based on colour. The terrain classification was performed using a Bayes classifier that used mixtures of Gaussians to model class likelihoods. The parameters of the mixture model were estimated from training data through the EM algorithm. All of these systems have met with some success in addressing a very difficult problem.
This essay has been submitted by a student. This is not an example of the work written by our professional essay writers. You can order our professional work here.