Target recognition in data from photon counting laser systems
Publish date: 2018-01-16
Report number: FOI-R--4479--SE
Pages: 32
Written in: Swedish
Keywords:
- Time correlated single photon counting
- ATR
- target detection
- target recognition
Abstract
This report forms a basis for decisions concerning research on signal processing for data from Time-Correlated Single-Photon Counting (TCSPC) laser systems. The report identifies, clarifies and gives suggestions on a number of areas considered particularly important in order to obtain good results concerning target detection and recognition. Equipped with appropriate optics, TCSPC systems allow for pixel sizes of about 5- 10 cm over 2 km measurement ranges, which is a prerequisite for applications such as long range target recognition. Complicating factors are the relatively small detector sizes (typically 128×32 pixels) compared to passive EO cameras and the degradation of data quality caused by turbulence when measuring long ranges close to the ground. In order to get a total data accuracy about 5-10 cm over several km, the position and orientation of the system have to be determined with very high accuracy, which requires an external positioning system and accurate methods for 3D data registration. The report gives an overview of novel methods related to automatic detection and recognition of partly occluded targets. In order to successfully solve such tasks, new approaches for target detection and machine learning techniques have to be considered. The main conclusions are that development in five areas is needed in order for FOI to be able to develop and assess robust techniques for target detection and recognition: 1. Sensor fusion with SLAM (Simultaneous Localization and Mapping) framework to provide an accurate estimation of the position and orientation of the sensor 2. Registration of TCSPC data for accurately matching parts of the collected 3D data with each other and transforming data to obtain maximal accuracy. 3. Development of new algorithms for target detection that can quickly determine what parts of the data could contain a target 4. Development of algorithms for target recognition based on local properties and machine learning. That gives the possibility to recognize partly occluded targets and targets affected by turbulence. 5. Scenario simulations for simulating targets under occlusion and under the influence of turbulence. Simulated data are to be used as a complement to real sensor data to support development, training and testing of algorithms. We suggest that already ongoing work for areas 1 and 2 continues along the same lines as today. For the third bullet results from other projects could be incorporated and directly used in this project. Concerning item 4, we suggest that initial studies of machine learning with Deep Learning techniques are initiated in the form of a Master Thesis work. The fifth area should be studied further within this project.