A highly robust vehicle collision detection method

Abstract

A highly robust vehicle collision detection method includes the following steps: optimizing the shape parameters a and b of the kernel function; inputting video frames and performing noise reduction preprocessing with the 1Shot-MaxPol algorithm; obtaining the optical flow field {ν xi , ν yj} using the Farneback algorithm; calculating the interaction field F(x, y) using the formula #imgabs0#; accumulating the area difference S t between the positive and negative regions of F(x, y), and obtaining the outlier #imgabs1# through quadratic regression. When #imgabs2#, it is determined that a collision has occurred.

Description

Technical Field

[0001] The invention belongs to the field of machine vision, and specifically uses road vehicle videos to detect whether a collision occurs between vehicles. Background Art

[0002] Traffic injuries have become one of the world's biggest public health threats. According to the World Health Organization, 1.24 million people die from traffic injuries each year, and another 50 million are injured. Many casualties in traffic accidents are caused by the fact that the injured do not receive timely rescue. The development of vehicle collision detection based on video surveillance is of great significance.

[0003] Video-based vehicle collision detection is one of the biggest challenges in the field of intelligent transportation. Traditional methods require target segmentation and tracking to obtain collision information, which has high computational complexity.

[0004] Kimin Yun et al. were inspired by the shape of the water surface when multiple objects move in the water, and used Gaussian functions to simulate water waves. This model was applied to all pixels of the image to obtain the motion interaction field. Since this method requires the use of a highly complex optical flow algorithm, it is difficult to use in real-time systems; and the motion interaction field is not ideal, making it difficult to stably use in different scenarios, especially nighttime scenarios with low signal-to-noise ratios. Summary of the invention

[0005] To overcome the above problems, a highly robust vehicle collision detection method according to the present invention comprises the following steps:

[0006] Step 1: Collect normal frame images and use differential evolution algorithm to optimize kernel function shape parameters a and b;

[0007] Step 2: Input the video frame and perform noise reduction preprocessing using the 1Shot-MaxPol algorithm;

[0008] Step 3: Use Farneback algorithm to get the optical flow field {ν xi ,ν yj};

[0009] Step 4: Use the formula Calculate the interaction field F(x,y), where the image size is M×N pixels, ω ij is the weight of each pixel, K(x,y,a,b;ν xi ,ν yj )for:

[0010] K(x,y,a,b;ν xi ,ν yj )=κ(x,y,a,b;x i +ν xi ,yj +ν yj )-κ(x,y,a,b;x i -ν xi ,y j -ν yj ),

[0011] Among them, κ(x,y,a,b; x c ,y c ) is a generalized bell-shaped kernel function with the center at (x c ,y c ), whose mathematical formula is:

[0012]

[0013] Among them, a and b are its two shape parameters;

[0014] Step 5: Add the area difference between the positive and negative regions of F(x,y) to get S t , get outliers through quadratic regression

[0015]

[0016] Step 6: When It is determined that a collision has occurred. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is the flow chart of this patent;

[0018] Figure 2 Real video frames at night;

[0019] Figure 3 Interaction fields obtained from real nighttime video frames;

[0020] Figure 4 Outliers at different moments of actual video frames at night;

[0021] Figure 5 Processed night video frame;

[0022] Figure 6 ideal interaction field;

[0023] Figure 7 This patent obtains abnormal values ​​at different moments;

[0024] Figure 8 The method described in the patent calculates time statistics for each part. DETAILED DESCRIPTION

[0025] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0026] The use of this patent Figure 1 The process shown processes a night-time video surveillance data.

[0027] Using differential evolution algorithm (population size NP = 20, mutation operator F = 0.5, crossover operator CR = 0.9, maximum number of evolution generations G = 50), collect normal frame image I0(t), according to the formula:

[0028]

[0029] Calculate the interaction field F(x,y) and add the area difference between the positive and negative regions of F(x,y) to get S t , the area difference S t As the objective function, the differential evolution algorithm is used to optimize the kernel function shape parameters a and b.

[0030] Real frames collected at night are as follows Figure 2 As shown, the interaction field is directly obtained as Figure 3 The corresponding outliers are shown in Figure 4 As shown in the figure, it can be seen that due to the poor video quality, there are two moments when the outliers exceed 0.7, which will cause misjudgment.

[0031] The video frames of nighttime video surveillance after noise reduction by the 1Shot-MaxPol algorithm are as follows: Figure 5 As shown, the interaction field is obtained as Figure 6 The corresponding outliers are shown in Figure 7 As shown, the threshold S th Taking 0.7 can accurately determine whether the vehicle has collided.

[0032] The algorithm of the present invention consists of four parts: video frame preprocessing, optical flow calculation, motion modulation field generation, and outlier calculation. The time consumed by each part is calculated in the PC platform window10 and MATLAB R2017b environment. Figure 8 As shown, it takes about 240ms in total to complete one algorithm.

[0033] The remarkable effects of the present invention are: it does not include vehicle target detection and tracking links, can effectively avoid environmental changes, accurately measure data, have high detection accuracy, and has a simple structure, low cost, easy operation, and can achieve long-term online monitoring.

Claims

1. A highly robust vehicle collision detection method, characterized in that: The following steps are involved: Step 1: Collect normal frame images and use differential evolution algorithm to optimize kernel function shape parameters a and b; Step 2: Input the video frame and perform noise reduction using the 1Shot-MaxPol algorithm; Step 3: Use Farneback algorithm to get the optical flow field {ν xi ,ν yj }; Step 4: Use the formula Calculate the interaction field F(x,y), where the image size is M×N pixels, ω ij is the weight of each pixel, K(x,y,a,b;ν xi ,ν yj )for: K(x,y,a,b;ν xi ,n yj )=κ(x,y,a,b;x i +n xi ,y j +n yj )-κ(x,y,a,b;x i -n xi ,y j -n yj ), Among them, κ(x,y,a,b; x c ,y c ) is a generalized bell-shaped kernel function with the center at (x c ,y c ), whose mathematical formula is: Among them, a and b are its two shape parameters; Step 5: Add the area difference between the positive and negative regions of F(x,y) to get S t , get outliers through quadratic regression Step 6: When S t abn >S th , it is determined that a collision occurs.

Images