Scout Aerial has been working on an automated shark detection algorithm for the past twelve months and this report details some of the results that were achieved during testing. Due to the innovative approach to this project, Scout Aerial needed a lot of sample data in order to build a baseline algorithm. Once we have identified the data requirements, and have developed a machine learning tool for this algorithm, we can then build a solution that is scalable and suited to the shark management process. By undertaking a pilot project, we are able to collect the data required, identify the particular environments (type and state), build a base algorithm and understand the scale requirements.

The main focus has been on training the algorithm which is an extremely manual process. A team of four pilots, two aircraft and supporting hardware and software was sent to test the real-time system.

We also have been working on the hardware design so that it may be attached to any RPA and will be fit-for-purpose.  The hardware can be offered in multiple configurations or “as-required” (multispectral, RGB, NIR,) and we will collect different data samples in order to “train” the base algorithm in “layers” to make it more accurate over time.

In regards to developing an algorithm,  the required specification aids in the identification and reduction of subconscious biases. By using an algorithm, decision making becomes a more rational process.

In additional to making the process more rational, use of an algorithm will make the process more efficient and more consistent. Efficiency is an inherent result of the analysis and specification process. Consistency comes from both the use of the same specified process and increased skill in applying the process. An algorithm develops and becomes “self-learning” over time, meaning a constant improvement in efficiency. The very first step to developing an algorithm is the collection of data that will be used as the foundation to build it.

Step 1a: Select dataset for training, historical and actual

Step 1b: The smaller part of the dataset is considered for testing.

Step 2a: The training dataset is passed to an ML algorithm like Linear Regression.

Step 2b: The testing dataset is populated and kept ready for evaluating.

Step 3a: The algorithm is applied to each data point from the training dataset.

Step 3b: The parameters are then applied to the test dataset to compare the outcome.

Step 3c: This process is iterated till the algorithm generates values that are a close match to the test data.

Step 4: The final model is evolved with the right set of parameters tuned for the given dataset.

Step 5: The model is used in production to predict based on the new data points.

The data output can be in many forms. We have chosen SMS with an Alert and geolocation, this can be attached to a KML and opened immediately in Google Earth. This can easily be pushed into an App if required. See below an example:

Get in contact with us today if you are interested in getting an algorithm or computer vision system developed to reduce your workload. Once developed, these tools can increase your efficiency drastically. Using technology to create change for the better, get onboard!

The identification of animals and objects in complex environments is an area of computer vision which has largely been neglected due to high expense of datasets and intensive computational requirements for analysis.  The challenges of identifying sharks and aquatic animals in environments littered with surfers, rocks and various other possible false positives required the development of a system incorporating the newest forms of computer vision machine learning. Simple methods such as color separation, intensity histograms and optical flow are all thrown off by the ever changing complex environment of the ocean.  For this reason a machine learning method that gradually improves with larger and larger datasets was the best solution.

The development of this system needs to overcome a few large obstacles that are present in this dynamic environment

• Ability to identify sharks from a minimum 60 meters altitude from a fast moving drone
• Detection through the highly changing environment of waves, glare and shadows created from the changing ocean
• False positives from surfers and a range of other marine life
• Ability to be run on an onboard single board computer

With use of machine learning methods the ability to essentially train out specific problematic areas and accentuate features of the shark for easier detection become feasible.  This has required the use of immensely huge datasets with over 50,000 individual images being utilized in order to successfully differentiate between the aggressive ocean and the targeted entity.  The result has been extremely successful over other techniques with low false positives and high identification in extremely complicated environments. The problems of waves, shadows and rocks have all been slowly trained out to a point where blind runs of unseen datasets produce high identification rates and low false positives.

The capabilities of this machine learning technique are near endless with the addition of more computational power the ability to run identification of multiple animals and objects can be undertaken. This can lead to a huge new range of possibilities in creating specific scenarios for warning systems and data collection.

Do you have a similar project in mind? Want to successfully integrate remotely piloted aircraft into your operation?  We would love to help you turn your idea into practical results. Contact us today.