The Battle Field Sensing Radar (BFSR) is a fully coherent Pulse-Doppler radar. It works at a very low peak power of 5W that makes it difficult to detect by enemy sensors. The radar operates over 21 channels in J band (10 to 18/20 GHz). It can be operated in all weather conditions, and during day and night. It is portable and can be mounted on vehciles.
Presently, the doppler sounds generated from battlefield sensing radar (BFSR) are manually classified. The human resources that classify these signals are trained over a significant period of time to be able to distinguish the signals generated by various categories of targets. In general, these categories are 1. Crawling man, 2. Group of men, 3. Light vehicle and 4. Heavy vehicle.
The Doppler sound effects from above 4 sources is significantly different. We use machine learning algorithms to train a model to be able to detect these differences and reasonably accurately predict the source. We use convoluted neural network (CNN) implementation of tensorflow, which is an open source software, supported by google to implement algorithms that use neural networks to train models. Current online resources show that tensorflow is better on several fronts compared to its peers. The research is readily available on internet.
The training process involves collecting the data from different radar setups. Each of such data collected is sanitized, this involves removing background noises, removing silence and then clipping them to appropriate size. The sanitized wave files are then transformed to spectrograms using fast Fourier transform. Thus, sanitized data is then passed to a custom Tensorflow based training process. A significant amount of data is passed thru training module while keeping a small percent for testing. After this process, a model is generated along with a confusion matrix that shows how well the training happened.
Fig 1. Sample FFT generated spectrograph of doppler sound spanning 1 second, first one is from group of man, second is from a light vehicle.
For predicting, the real time data is captured from the radars audio port. This data is also sanitized and clipped to appropriate size. The predictor with the model then classifies the input signal into one of the categories with a confidence value. This value can help tune the sensitivity of the prediction algorithm.
The two phases are depicted in the diagram below.
Fig 2. Shows the logical flow of events for the two phases 1. Training and 2. Prediction.
It is important that false positives are utterly avoided. With propriotory techniques we have been able to reduce false positives to less than 1%.
As we speak the data is being collected across various terrains to improve the target classification.
Technical details have been suppressed in this article, to know more contact firstname.lastname@example.org.