Bloodstain impact stains are typically analysed using the well-known ellipse model. This model, however, simplifies the process of stain formation to a projection of spherical droplet. HemoVision does not use the ellipse model, but instead uses a statistical shape model that was trained on images of impact stains with known impact angles. This model can be fitted automatically to new stain images, and has a higher accuracy compared to automatic ellipse fitting.

To train our statistical shape model, we first created a dataset of impact stains with know impact angles by dropping pigs blood onto plain white paper with impact angles ranging from 5 to 90 degrees (with 2.5 degree increments). The bloodstains were photographed and subsequently processed to extract their contours. This processing involved segmentation, (semi-)automatic tail removal, and contour resampling, resulting in 100 homologous landmarks for each of the 367 stains.

Once the contours were extracted, we aligned them to remove all but the shape information, used a Principal Component Analysis (PCA) on the landmark coordinates to train our model. Just one principal component was retained, since it accounted for a total of 97.7% of variation in our dataset. The resulting 1D principal component scores were then linked to impact angles using a 3^rd order polynomial regression model.

Our statistical shape model was combined with an automatic fitting algorithm, which was then used to analyse each stain in our dataset using a leave-one-out procedure. Estimated impact angles were compared to the true values. For impact angles between 5 and 60 degrees, our method achieved a root-mean-square error of 1.44 degrees compared to 2.09 degrees for automatic ellipse fitting.

For a more details about our statistical shape model or its evaluation, please refer to the full article here.

Based on: Joris, Philip, et al. "Calculation of bloodstain impact angles using an Active Bloodstain Shape Model." Journal of Forensic Radiology and Imaging 2.4 (2014): 188-198.