Fighting Poaching through High Precision Real-Time Gunshot Detection Using Deep Learning

Описание: Poaching remains a global issue despite increased surveillance and collaborative efforts. The advancement of low-cost Autonomous Recording Units has enabled acoustic monitoring over large spatial and temporal scales. However, high false-positive rates and large file sizes of existing classifiers have rendered real-time acoustic detection of poaching impractical. I propose a low-footprint acoustic gunshot detection model and a cross-device communication pipeline to reduce false positives, suitable for on-device, real-time rainforest gunshot detection.A custom deep neural network consisting of 2D and 1D convolution layers followed by a Long-Short Term Memory layer was trained on open-source rainforest|gunshot audio from Belize, with depthwise-separable convolutions to reduce size. A novel approach to reduce false positives is the use of cross-device communication (CDC). I propose the Sensor Analysis and Integration Layer (SAIL), a custom probability weighting function designed to leverage information from multiple sensors and map three or more predictions into a final prediction. CDC was simulated by performing three sound-attenuation augmentations on every waveform in a test dataset of 7190 files, emulating a three-sensor array for each waveform. The proposed gunshot detection network produced confidence probabilities for each augmented waveform, which were then fed to the SAIL function. To evaluate SAIL under varying sensor placements, three simulation configurations were created, each increasing in augmentation, or sensor array, variance, and ten simulations were run per configuration, totaling thirty. SAIL achieved a statistically significant average FPR reduction of 29%, from 0.126 to 0.09, compared to non-simulation classification. Additionally, the proposed network achieved a validation F1 of .89 at 598KB; a competitive performance at 1|50th the size of leading classifiers. Combined with the significant FPR reduction across all simulations, SAIL and efficient deep learning present great promise in addressing the major obstacles currently preventing acoustic monitoring from becoming a cost-effective solution for long-term, large-scale, real-time poaching detection.