drone-killer

Drone detection system using an SSD MobileNet v1 model converted to TensorFlow Lite. Processes video files or live camera feeds, draws bounding boxes around detected drones, and optionally saves the annotated output.

Usage

./demo <input> [output_path] [--server-url URL]

Input type	Argument example	Description
Video file	./demo video.mp4 output.mp4	Process a video file and save annotated output
Camera index	./demo 0	Open /dev/video0 (V4L2 / USB camera)
Camera + record	./demo 0 output.mp4	Live camera with recording to file
Camera + event upload	./demo 0 --server-url http://localhost:8001/events	Save/POST detection snippets to backend events API
GStreamer pipeline	./demo "libcamerasrc ! video/x-raw,width=640,height=480,framerate=30/1 ! videoconvert ! appsink"	Pi libcamera stack (Bullseye+)

• Video file mode: processes all frames, writes output, and exits.
• Live camera mode: displays a window with detections in real-time. Press q to quit. Output file is optional.
• Event upload mode: when --server-url is set, each throttled detection snapshot is uploaded as multipart form data with metadata and file field snippet.

Build (native, on ARM64 Debian 12 or 13 VM)

Prerequisites

sudo apt update
sudo apt install -y cmake g++ git wget unzip pkg-config \
    libopencv-dev \
    libavcodec-dev libavformat-dev libswscale-dev \
    libcurl4-openssl-dev

Build TFLite C library (same as image-comp-installer.sh)

git clone --depth 1 --branch v2.18.0 https://github.com/tensorflow/tensorflow.git /tmp/tf
cd /tmp/tf && mkdir build && cd build
cmake ../tensorflow/lite/c \
    -DCMAKE_BUILD_TYPE=Release \
    -DTFLITE_ENABLE_XNNPACK=OFF
make -j$(nproc)

sudo find . -name 'libtensorflowlite_c*' -type f -exec cp {} /usr/local/lib/ \;
cd /tmp/tf
sudo find tensorflow/lite -name '*.h' -exec sh -c \
    'mkdir -p "/usr/local/include/$(dirname "$1")" && cp "$1" "/usr/local/include/$1"' _ {} \;
sudo ldconfig
rm -rf /tmp/tf

Build the project

cd ~/drone-killer
cmake -B build -DCMAKE_BUILD_TYPE=Release -DTFLITE_DIR=/usr/local
cmake --build build

Test

./build/demo V_DRONE_024.mp4 output.mp4

Deploy to Raspberry Pi

Copy the following to the Pi:

build/demo                        # binary
model/drone_detect.tflite         # TFLite model (no Flex ops)
model/anchors.bin                 # precomputed anchor boxes
model/object-detection.pbtxt      # label map

On the Pi, install runtime dependencies:

sudo apt install libopencv-dev libavcodec-dev libavformat-dev libswscale-dev

Run:

# From a video file
./demo video.mp4 output.mp4

# From USB camera
./demo 0

# From Pi camera (libcamera)
./build/demo "libcamerasrc ! videoconvert ! video/x-raw,format=BGR ! appsink drop=true max-buffers=1 sync=false" --server-url http://2doorspacemachine.local:8001/events

Model

SSD MobileNet v1 trained for drone detection. The TFLite model outputs raw box encodings and class logits — post-processing (box decoding, sigmoid, NMS) is done in C++.

• Input: 300x300x3 float32, normalized to [-1, 1]
• Output 0: raw box encodings [1, 1917, 4]
• Output 1: raw class scores [1, 1917, 2] (background + drone)