تجزیه و تحلیل عملکرد از یک الگوریتم ساده تشخیص خودرو

We have performed an end-to-end analysis of a simple model-based vehicle detection algorithm for aerial parking lot images. We constructed a vehicle detection operator by combining four elongated edge operators designed to collect edge responses from the sides of a vehicle. We derived the detection and localization performance of this algorithm, and verified them by experiments. Performance degradation due to different camera angles and illuminations was also examined using simulated images. Another important aspect of performance characterization — whether and how much prior information about the scene improves performance — was also investigated. As a statistical diagnostic tool for the detection performance, a computational approach employing bootstrap was used.

The task of a computer vision algorithm can be specified in terms of two components: the range of images to be processed and the performance criterion that the algorithm should try to achieve. The algorithm can then be designed to handle every image in the given class, and at the same time, to optimize the specified criterion function. The performance of the algorithm is evaluated in terms of these components, using two different methodologies: theoretical formulation of the relation between the imaging conditions and the criterion function, and empirical evaluation using real or simulated imagery. If there is an underlying model for the image and for the possible perturbations of the image that we may expect in a real environment, and also if the algorithm is based on some mathematical formulation, it may be possible to predict the performance theoretically. Haralick [14] has asserted that performance characterization of a vision algorithm has to do with establishing the random variations in the output data that result from the random variations and imperfections in the input data. However, a scene model is not always available or is usually not complete; hence, the error propagation computed using a theoretical model is usually a crude approximation. Numerous imaging conditions can affect the performance of an algorithm, including scene noise and camera and lighting angles. If we can quantify these scene factors and if we have a very large set of images annotated with these parameters, experiments on this set of images would give a very informative performance measure for the algorithm. In reality, gathering data for such a complete and systematic evaluation is not usually possible. Most algorithms are based on some model or an underlying theoretical framework; however, analytical tools for predicting the behavior of algorithms are not usually available. On the other hand, it is usually easy to generate simulated images and investigate algorithm performance using these images. If a perturbation model for the scene is available, we can generate perturbed images and conduct a systematic evaluation. While this kind of experiment is commonplace in computer vision, there have been objections to this method: the scene model as well as the perturbation model are far from realistic. This is true; however, as stated in Ref. [11], simulation is inevitable to verify the correctness of implementations, and also to analyze the behavior of algorithms under varying conditions for which real images are not available. Some scene environments are not easy to model, and it is usually impossible to acquire annotated ground-truth data for all the relevant sets of parameters. In our analysis of the vehicle detection task, we followed the above guidelines. The three most important characterizations of the task of vehicle detection are the detection, false alarm, and localization performances. There are numerous relevant scene factors, including resolution, noise, camera and lighting conditions, parking lot layout, etc. It is important to identify which parameters affect the criterion function significantly and which do not. We have found that the camera and illumination angles are the most crucial factors. We tested extreme values of these factors to verify the degradation of performance. With low illumination angles, the vehicle detector produces many false alarms due to long shadows; also, very oblique camera angles contribute to a poor detection rate. Since these angles are the most crucial factors in algorithm performance, we can conjecture that some prior information about these parameters can reduce errors. The noise level of the image how the vehicles are positioned in the parking lot, and the colors of the vehicles also affect performance. We have designed a vehicle detector based on a simple vehicle model and Canny's formulation of edge detection, and have investigated these issues in a controlled environment using a parking lot scene simulator, as well as in a real environment. The utility of site information was investigated using real images with ground-truthed vehicles and annotated parking lot orientations. By using the local linearity of the vehicle operator response, statistical properties of its detection and localization performance were derived, and they were then verified by simulation. We can view this formulation as error propagation: via the geometry of the vehicle model, from the image domain to the detection probability and parameter space. This framework is one of the contributions of this work, which can be extended to more general problems. As it turned out, the most serious problem in vehicle detection is the occurrence of many false alarms, due to spurious responses from adjacent vehicles, road structures, and buildings. We developed an empirical hypothesis test to remove false alarms and self-diagnose the detection and localization performance. This paper is organized as follows: Section 2 introduces the task of vehicle detection (Section 2.1) along with a review of related work (Section 2.2). By extending Canny's formulation of step edge detection (Section 2.3), we provide a solution to the problem of vehicle detection (Section 2.4). The output of the algorithm for a small parking lot image is discussed in Section 2.5. A mathematical formulation is given in 3.1 and 3.2. Performance evaluation using simulated images is discussed in Section 4. Experiments on real ground-truthed images are given in Section 5. Finally, a method of self-diagnosing the performance using bootstrap is introduced in Section 6.

The performance of a simple model-based vehicle detection algorithm under a wide range of operating environments has been investigated using mathematical analysis and empirical evaluation. It was verified that low illumination and/or acquisition angle contribute to poor detection and localization, and that site information can improve them. We derived statistical properties of the detection and localization performances in the presence of noise. The trade-off between detection and localization performances can be exploited to meet the requirements of given applications. Since our vehicle operator can be generalized for detecting shapes other than rectangles, this framework can be directly applied to more general shape detection problems. Through our analysis, the use of a more sophisticated vehicle model was suggested for performance improvement in adverse environments. Statistical validation using bootstrap was found to be effective in removing false alarms. However, since this method also depends on the vehicle model, further improvement can be made by using more sophisticated models.