Copilot : Driving assistance on mobile devices

Lane and obstacle detection for active assistance during driving.

Vehicle Position + collision time superposed in the top view

Accompanying article https://towardsdatascience.com/copilot-driving-assistance-635e1a50f14

Global annual road accidents fatalities total about 1.5 million which is just about the population of Mauritius. 90% of these occur in low and middle income countries which have less than half of the total vehicles in the world. Advanced driver-assistance systems (ADAS) Lane detection, collision warning are present in less than 0.1% of the vehicles. They are almost non existent in developing countries. Median Smartphone ownership in emerging economies is about 10 times as high as that of four wheeler. While we already have semi autonomous vehicles running about in parts of the world. This repository checks how close we might come to using a mobile computing platform as an ADAS copilot.

DOWNLOAD WEIGHTS AND CODE

! git clone https://github.com/visualbuffer/copilot.git
! mv copilot/* ./
! wget -P ./model_data/ https://s3-ap-southeast-1.amazonaws.com/deeplearning-mat/backend.h5

Robustness for different illumination conditionsz

USAGE EXAMPLE

from frame import FRAME

file_path =  "videos/highway.mp4"# <== Upload appropriate file          
video_out = "videos/output11.mov"
frame =  FRAME( 
    ego_vehicle_offset = .15,                       # SELF VEHICLE OFFSET
    yellow_lower = np.uint8([ 20, 50,   100]),      # LOWER YELLOW HLS THRESHOLD
    yellow_upper = np.uint8([35, 255, 255]),        # UPER YELLOW HLS THRESHOLD
    white_lower = np.uint8([ 0, 200,   0]),         # LOWER WHITE THRESHOLD
    white_upper = np.uint8([180, 255, 100]),        # UPPER WHITE THRESHOLD
    lum_factor = 118,                               # NORMALIZING LUM FACTOR
    max_gap_th = 0.45,                              # MAX GAP THRESHOLD
    YOLO_PERIOD = .25,                              # YOLO PERIOD
    lane_start=[0.35,0.75] ,                        # LANE INITIATION
    verbose = 3)                                    # VERBOSITY
frame.process_video(file_path, 1,\
        video_out = video_out,pers_frame_time =144,\
        t0  =144 , t1 =150)#None)

PARAMETER	Description
SELF VEHICLE OFFSET	Trim off from bottom edge video if ego vehicle covers part of the frame % of front view
LOWER YELLOW HLS THRESHOLD	Lower yellow HLS threshold used to prepare the mask. Tune down if yellow lane is not detected, up if all the foilage is
UPPER YELLOW HLS THRESHOLD	Upper threshold for identifying yellow lanes
LOWER WHITE THRESHOLD	Lower yellow HLS threshold used to prepare the mask. Tune up saturation if foilage lights up the entire scene
UPPER WHITE THRESHOLD
NORMALIZING LUM FACTOR	Factor used to normalize luminosity against, reducing increses lower Lum threshold
MAX GAP THRESHOLD	Max continous gap tollerated in the lane detection % of top-view height
YOLO PERIOD	Period [s] after which YOLO is detected, typ 2s reducing decreases processing fps increases detection
LANE INITIATION	intial guess for lane start % of top-view width
VERBOSITY	1 Show lesser,2 Show less,3 Show everything

Detecting lane change automatically

Notebooks

DIRECTORY	COLAB
./notebooks/coPilot.ipynb	https://colab.research.google.com/drive/1CdqDXZqssDgSC35W4A-4Gp8kfqzyPKug

Ref:

https://github.com/qqwweee/keras-yolo3